diff --git a/agent/bin/agent.dart b/agent/bin/agent.dart
index d69b1c9b41..72fb461f3c 100644
--- a/agent/bin/agent.dart
+++ b/agent/bin/agent.dart
@@ -10,7 +10,6 @@ import 'package:http/http.dart';
 
 import 'package:cocoon_agent/src/agent.dart';
 import 'package:cocoon_agent/src/commands/ci.dart';
-import 'package:cocoon_agent/src/commands/run.dart';
 import 'package:cocoon_agent/src/utils.dart';
 
 Future<Null> main(List<String> rawArgs) async {
@@ -26,7 +25,6 @@ Future<Null> main(List<String> rawArgs) async {
       defaultsTo: '.'
     );
   argParser.addCommand('ci');
-  argParser.addCommand('run', RunCommand.argParser);
 
   ArgResults args = argParser.parse(rawArgs);
 
@@ -45,7 +43,6 @@ Future<Null> main(List<String> rawArgs) async {
   }
 
   registerCommand(new ContinuousIntegrationCommand(agent));
-  registerCommand(new RunCommand(agent));
 
   if (args.command == null) {
     print('No command specified, expected one of: ${allCommands.keys.join(', ')}');
diff --git a/agent/lib/src/adb.dart b/agent/lib/src/adb.dart
index 6188ce6db6..0bc4be27b6 100644
--- a/agent/lib/src/adb.dart
+++ b/agent/lib/src/adb.dart
@@ -3,7 +3,6 @@
 // found in the LICENSE file.
 
 import 'dart:async';
-import 'dart:math' as math;
 
 import 'package:meta/meta.dart';
 
@@ -28,24 +27,8 @@ abstract class DeviceDiscovery {
     }
   }
 
-  /// Selects a device to work with, load-balancing between devices if more than
-  /// one are available.
-  ///
-  /// A task performing multiple actions on a device would call this method
-  /// first then follow up by calling [workingDevice]. Calling [workingDevice]
-  /// repeatedly must guarantee to return the same device.
-  Future<Null> chooseWorkingDevice();
-
-  /// Returns current working device.
-  ///
-  /// Must be called _after_ a successful call to [chooseWorkingDevice]. May be
-  /// called repeatedly to perform multiple operations on one specific device;
-  /// guarantees to return the same device until the next time
-  /// [chooseWorkingDevice] is called.
-  Device get workingDevice;
-
   /// Lists all available devices' IDs.
-  Future<List<String>> discoverDevices();
+  Future<List<Device>> discoverDevices();
 
   /// Checks the health of the available devices.
   Future<Map<String, HealthCheckResult>> checkDevices();
@@ -94,37 +77,8 @@ class AndroidDeviceDiscovery implements DeviceDiscovery {
 
   AndroidDeviceDiscovery._();
 
-  AndroidDevice _workingDevice;
-
-  @override
-  AndroidDevice get workingDevice {
-    if (_workingDevice == null) {
-      throw new StateError(
-        'No working device chosen. Call `chooseWorkingDevice` prior to calling '
-        'the `workingDevice` getter.',
-      );
-    }
-
-    return _workingDevice;
-  }
-
-  /// Picks a random Android device out of connected devices and sets it as
-  /// [workingDevice].
   @override
-  Future<Null> chooseWorkingDevice() async {
-    List<Device> allDevices = (await discoverDevices())
-      .map((String id) => new AndroidDevice(deviceId: id))
-      .toList();
-
-    if (allDevices.isEmpty)
-      throw 'No Android devices detected';
-
-    // TODO(yjbanov): filter out and warn about those with low battery level
-    _workingDevice = allDevices[new math.Random().nextInt(allDevices.length)];
-  }
-
-  @override
-  Future<List<String>> discoverDevices() async {
+  Future<List<Device>> discoverDevices() async {
     List<String> output = (await eval(config.adbPath, ['devices', '-l'], canFail: false))
         .trim().split('\n');
     List<String> results = <String>[];
@@ -150,21 +104,22 @@ class AndroidDeviceDiscovery implements DeviceDiscovery {
       }
     }
 
-    return results;
+    return results
+      .map((String id) => new AndroidDevice(deviceId: id))
+      .toList();
   }
 
   @override
   Future<Map<String, HealthCheckResult>> checkDevices() async {
     Map<String, HealthCheckResult> results = <String, HealthCheckResult>{};
-    for (String deviceId in await discoverDevices()) {
+    for (AndroidDevice device in await discoverDevices()) {
       try {
-        AndroidDevice device = new AndroidDevice(deviceId: deviceId);
         // Just a smoke test that we can read wakefulness state
         // TODO(yjbanov): check battery level
         await device._getWakefulness();
-        results['android-device-$deviceId'] = new HealthCheckResult.success();
+        results['android-device-${device.deviceId}'] = new HealthCheckResult.success();
       } catch(e, s) {
-        results['android-device-$deviceId'] = new HealthCheckResult.error(e, s);
+        results['android-device-${device.deviceId}'] = new HealthCheckResult.error(e, s);
       }
     }
     return results;
@@ -260,37 +215,8 @@ class IosDeviceDiscovery implements DeviceDiscovery {
 
   IosDeviceDiscovery._();
 
-  IosDevice _workingDevice;
-
   @override
-  IosDevice get workingDevice {
-    if (_workingDevice == null) {
-      throw new StateError(
-        'No working device chosen. Call `chooseWorkingDevice` prior to calling '
-        'the `workingDevice` getter.',
-      );
-    }
-
-    return _workingDevice;
-  }
-
-  /// Picks a random iOS device out of connected devices and sets it as
-  /// [workingDevice].
-  @override
-  Future<Null> chooseWorkingDevice() async {
-    List<IosDevice> allDevices = (await discoverDevices())
-      .map((String id) => new IosDevice(deviceId: id))
-      .toList();
-
-    if (allDevices.length == 0)
-      throw 'No iOS devices detected';
-
-    // TODO(yjbanov): filter out and warn about those with low battery level
-    _workingDevice = allDevices[new math.Random().nextInt(allDevices.length)];
-  }
-
-  @override
-  Future<List<String>> discoverDevices() async {
+  Future<List<Device>> discoverDevices() async {
     // TODO: use the -k UniqueDeviceID option, which requires much less parsing.
     List<String> iosDeviceIds = grep('UniqueDeviceID', from: await eval('ideviceinfo', []))
       .map((String line) => line.split(' ').last).toList();
@@ -298,15 +224,17 @@ class IosDeviceDiscovery implements DeviceDiscovery {
     if (iosDeviceIds.isEmpty)
       throw 'No connected iOS devices found.';
 
-    return iosDeviceIds;
+    return iosDeviceIds
+      .map((String id) => new IosDevice(deviceId: id))
+      .toList();
   }
 
   @override
   Future<Map<String, HealthCheckResult>> checkDevices() async {
     Map<String, HealthCheckResult> results = <String, HealthCheckResult>{};
-    for (String deviceId in await discoverDevices()) {
+    for (Device device in await discoverDevices()) {
       // TODO: do a more meaningful connectivity check than just recording the ID
-      results['ios-device-$deviceId'] = new HealthCheckResult.success();
+      results['ios-device-${device.deviceId}'] = new HealthCheckResult.success();
     }
     return results;
   }
diff --git a/agent/lib/src/agent.dart b/agent/lib/src/agent.dart
index 0aafb91d04..67fb4f25bf 100644
--- a/agent/lib/src/agent.dart
+++ b/agent/lib/src/agent.dart
@@ -9,15 +9,6 @@ import 'package:args/args.dart';
 import 'package:http/http.dart';
 import 'package:meta/meta.dart';
 
-import 'package:cocoon_agent/src/adb.dart';
-import 'package:cocoon_agent/src/analysis.dart';
-import 'package:cocoon_agent/src/framework.dart';
-import 'package:cocoon_agent/src/gallery.dart';
-import 'package:cocoon_agent/src/golem.dart';
-import 'package:cocoon_agent/src/perf_tests.dart';
-import 'package:cocoon_agent/src/refresh.dart';
-import 'package:cocoon_agent/src/hot_dev_cycle.dart';
-import 'package:cocoon_agent/src/size_tests.dart';
 import 'package:cocoon_agent/src/utils.dart';
 
 /// Contains information about a Cocoon task.
@@ -49,25 +40,6 @@ class Agent {
     return JSON.decode(resp.body);
   }
 
-  Future<BuildResult> performTask(CocoonTask reservation) async {
-    int golemRevision = await computeGolemRevision();
-    Task task = await getTask(reservation);
-    TaskRunner runner = new TaskRunner(reservation.revision, golemRevision, <Task>[task]);
-    try {
-      return await runner.run();
-    } finally {
-      await _screenOff();
-    }
-  }
-
-  Future<Null> _screenOff() async {
-    try {
-      await devices.workingDevice.sendToSleep();
-    } catch(error, stackTrace) {
-      print('Failed to turn off screen: $error\n$stackTrace');
-    }
-  }
-
   Future<Null> uploadLogChunk(CocoonTask task, String chunk) async {
     String url = '$baseCocoonUrl/api/append-log?ownerKey=${task.key}';
     Response resp = await httpClient.post(url, body: chunk);
@@ -77,36 +49,6 @@ class Agent {
     }
   }
 
-  Future<Task> getTask(CocoonTask task) async {
-    DateTime revisionTimestamp = await getFlutterRepoCommitTimestamp(task.revision);
-    String dartSdkVersion = await getDartVersion();
-
-    List<Task> allTasks = <Task>[
-      createComplexLayoutScrollPerfTest(),
-      createComplexLayoutScrollPerfTest(ios: true),
-      createFlutterGalleryStartupTest(),
-      createFlutterGalleryStartupTest(ios: true),
-      createComplexLayoutStartupTest(),
-      createComplexLayoutStartupTest(ios: true),
-      createFlutterGalleryBuildTest(),
-      createComplexLayoutBuildTest(),
-      createGalleryTransitionTest(),
-      createGalleryTransitionTest(ios: true),
-      createBasicMaterialAppSizeTest(),
-      createAnalyzerCliTest(sdk: dartSdkVersion, commit: task.revision, timestamp: revisionTimestamp),
-      createAnalyzerServerTest(sdk: dartSdkVersion, commit: task.revision, timestamp: revisionTimestamp),
-      createRefreshTest(commit: task.revision, timestamp: revisionTimestamp),
-      createHotDevCycleTest(commit: task.revision, timestamp: revisionTimestamp),
-    ];
-
-    return allTasks.firstWhere(
-      (Task t) => t.name == task.name,
-      orElse: () {
-        throw 'Task $task.name not found';
-      }
-    );
-  }
-
   /// Reserves a task in Cocoon backend to be performed by this agent.
   ///
   /// If not tasks are available returns `null`.
diff --git a/agent/lib/src/analysis.dart b/agent/lib/src/analysis.dart
deleted file mode 100644
index 2440aeb12c..0000000000
--- a/agent/lib/src/analysis.dart
+++ /dev/null
@@ -1,107 +0,0 @@
-// Copyright (c) 2016 The Chromium Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style license that can be
-// found in the LICENSE file.
-
-import 'dart:async';
-import 'dart:io';
-
-import 'package:meta/meta.dart';
-import 'package:path/path.dart' as path;
-
-import 'benchmarks.dart';
-import 'framework.dart';
-import 'utils.dart';
-
-Task createAnalyzerCliTest({
-  @required String sdk,
-  @required String commit,
-  @required DateTime timestamp
-}) {
-  return new AnalyzerCliTask(sdk, commit, timestamp);
-}
-
-Task createAnalyzerServerTest({
-  @required String sdk,
-  @required String commit,
-  @required DateTime timestamp
-}) {
-  return new AnalyzerServerTask(sdk, commit, timestamp);
-}
-
-abstract class AnalyzerTask extends Task {
-  AnalyzerTask(String name) : super(name);
-
-  Benchmark benchmark;
-
-  @override
-  Future<TaskResultData> run() async {
-    section(benchmark.name);
-    await runBenchmark(benchmark, iterations: 3, warmUpBenchmark: true);
-    return benchmark.bestResult;
-  }
-}
-
-class AnalyzerCliTask extends AnalyzerTask {
-  AnalyzerCliTask(String sdk, String commit, DateTime timestamp) : super('analyzer_cli__analysis_time') {
-    this.benchmark = new FlutterAnalyzeBenchmark(sdk, commit, timestamp);
-  }
-}
-
-class AnalyzerServerTask extends AnalyzerTask {
-  AnalyzerServerTask(String sdk, String commit, DateTime timestamp) : super('analyzer_server__analysis_time') {
-    this.benchmark = new FlutterAnalyzeAppBenchmark(sdk, commit, timestamp);
-  }
-}
-
-class FlutterAnalyzeBenchmark extends Benchmark {
-  FlutterAnalyzeBenchmark(this.sdk, this.commit, this.timestamp)
-    : super('flutter analyze --flutter-repo');
-
-  final String sdk;
-  final String commit;
-  final DateTime timestamp;
-
-  File get benchmarkFile => file(path.join(config.flutterDirectory.path, 'analysis_benchmark.json'));
-
-  @override
-  TaskResultData get lastResult => new TaskResultData.fromFile(benchmarkFile);
-
-  @override
-  Future<num> run() async {
-    rm(benchmarkFile);
-    await inDirectory(config.flutterDirectory, () async {
-      await flutter('analyze', options: ['--flutter-repo', '--benchmark']);
-    });
-    return addBuildInfo(benchmarkFile, timestamp: timestamp, expected: 25.0, sdk: sdk, commit: commit);
-  }
-}
-
-class FlutterAnalyzeAppBenchmark extends Benchmark {
-  FlutterAnalyzeAppBenchmark(this.sdk, this.commit, this.timestamp)
-    : super('analysis server mega_gallery');
-
-  final String sdk;
-  final String commit;
-  final DateTime timestamp;
-
-  @override
-  TaskResultData get lastResult => new TaskResultData.fromFile(benchmarkFile);
-
-  Directory get megaDir => dir(path.join(config.flutterDirectory.path, 'dev/benchmarks/mega_gallery'));
-  File get benchmarkFile => file(path.join(megaDir.path, 'analysis_benchmark.json'));
-
-  Future<Null> init() {
-    return inDirectory(config.flutterDirectory, () async {
-      await dart(['dev/tools/mega_gallery.dart']);
-    });
-  }
-
-  @override
-  Future<num> run() async {
-    rm(benchmarkFile);
-    await inDirectory(megaDir, () async {
-      await flutter('watch', options: ['--benchmark']);
-    });
-    return addBuildInfo(benchmarkFile, timestamp: timestamp, expected: 10.0, sdk: sdk, commit: commit);
-  }
-}
diff --git a/agent/lib/src/benchmarks.dart b/agent/lib/src/benchmarks.dart
deleted file mode 100644
index 5877218633..0000000000
--- a/agent/lib/src/benchmarks.dart
+++ /dev/null
@@ -1,56 +0,0 @@
-// Copyright (c) 2016 The Chromium Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style license that can be
-// found in the LICENSE file.
-
-import 'dart:async';
-
-import 'framework.dart';
-
-abstract class Benchmark {
-  Benchmark(this.name);
-
-  final String name;
-
-  TaskResultData bestResult;
-
-  Future<Null> init() => new Future.value();
-
-  Future<num> run();
-  TaskResultData get lastResult;
-
-  String toString() => name;
-}
-
-Future<num> runBenchmark(Benchmark benchmark, {
-  int iterations: 1,
-  bool warmUpBenchmark: false
-}) async {
-  await benchmark.init();
-
-  List<num> allRuns = <num>[];
-
-  num minValue;
-
-  if (warmUpBenchmark)
-    await benchmark.run();
-
-  while (iterations > 0) {
-    iterations--;
-
-    print('');
-
-    try {
-      num result = await benchmark.run();
-      allRuns.add(result);
-
-      if (minValue == null || result < minValue) {
-        benchmark.bestResult = benchmark.lastResult;
-        minValue = result;
-      }
-    } catch (error) {
-      print('benchmark failed with error: $error');
-    }
-  }
-
-  return minValue;
-}
diff --git a/agent/lib/src/buildbot.dart b/agent/lib/src/buildbot.dart
deleted file mode 100644
index 44566fc131..0000000000
--- a/agent/lib/src/buildbot.dart
+++ /dev/null
@@ -1,141 +0,0 @@
-// Copyright 2016 The Chromium Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style license that can be
-// found in the LICENSE file.
-
-import 'dart:async';
-import 'dart:convert';
-import 'dart:io';
-
-import 'package:http/http.dart' as http;
-
-const String _jsonEndpoint = 'https://build.chromium.org/p/client.flutter/json';
-
-/// Looks up the latest `git` commit SHA that succeeded both on Linux and Mac
-/// buildbots.
-Future<String> getLatestGreenRevision() async {
-  BuilderClient linuxBuilder = new BuilderClient('Linux');
-  BuilderClient macBuilder = new BuilderClient('Mac');
-  List<int> linuxBuildNumbers = await linuxBuilder.listRecentBuilds();
-  List<int> macBuildNumbers = await macBuilder.listRecentBuilds();
-  Set<String> greenLinuxRevisions = new Set<String>();
-  Set<String> greenMacRevisions = new Set<String>();
-
-  // Keep fetching builds until we find a build that was green both on Mac
-  // and Linux buildbots.
-  while (greenLinuxRevisions.intersection(greenMacRevisions).isEmpty &&
-      linuxBuildNumbers.isNotEmpty &&
-      macBuildNumbers.isNotEmpty) {
-    BuildInfo linuxBuild = await linuxBuilder.getBuild(linuxBuildNumbers.removeLast());
-    if (linuxBuild.isGreen)
-      greenLinuxRevisions.add(linuxBuild.revision);
-
-    BuildInfo macBuild = await macBuilder.getBuild(macBuildNumbers.removeLast());
-    if (macBuild.isGreen)
-      greenMacRevisions.add(macBuild.revision);
-  }
-
-  Set<String> intersection = greenLinuxRevisions.intersection(greenMacRevisions);
-  if (intersection.isEmpty) {
-    // No builds that are green on both Mac and Linux
-    return null;
-  }
-
-  return intersection.single;
-}
-
-class BuildInfo {
-  BuildInfo(this.builderName, this.number, this.isGreen, this.revision);
-
-  final String builderName;
-  final int number;
-  final bool isGreen;
-  final String revision;
-}
-
-class BuilderClient {
-  BuilderClient(this.builderName);
-
-  final String builderName;
-
-  String get builderUrl => '${_jsonEndpoint}/builders/$builderName';
-
-  Future<BuildInfo> getBuild(int buildNumber) async {
-    Map<String, dynamic> buildJson = await _getJson('$builderUrl/builds/$buildNumber');
-
-    return new BuildInfo(
-      builderName,
-      buildNumber,
-      _isGreen(buildJson),
-      _getBuildProperty(buildJson, 'git_revision')
-    );
-  }
-
-  Future<List<int>> listRecentBuilds() async {
-    Map<String, dynamic> resp = await _getJson('$builderUrl/builds');
-    return resp.keys.map(int.parse).toList();
-  }
-}
-
-Future<dynamic> _getJson(String url) async {
-  return JSON.decode((await http.get(url)).body);
-}
-
-/// Properties are encoded as:
-///
-///     {
-///       "properties": [
-///         [
-///           "name1",
-///           value1,
-///           ... things we don't care about ...
-///         ],
-///         [
-///           "name2",
-///           value2,
-///           ... things we don't care about ...
-///         ]
-///       ]
-///     }
-dynamic _getBuildProperty(Map<String, dynamic> buildJson, String propertyName) {
-  List<List<dynamic>> properties = buildJson['properties'];
-  for (List<dynamic> property in properties) {
-    if (property[0] == propertyName)
-      return property[1];
-  }
-  return null;
-}
-
-/// Parses out whether the build was successful.
-///
-/// Successes are encoded like this:
-///
-///     "text": [
-///       "build",
-///       "successful"
-///     ]
-///
-/// Exceptions are encoded like this:
-///
-///     "text": [
-///       "exception",
-///       "steps",
-///       "exception",
-///       "flutter build apk material_gallery"
-///     ]
-///
-/// Errors are encoded like this:
-///
-///     "text": [
-///       "failed",
-///       "steps",
-///       "failed",
-///       "flutter build ios simulator stocks"
-///     ]
-bool _isGreen(Map<String, dynamic> buildJson) {
-  if (buildJson['text'] == null || buildJson['text'].length < 2) {
-    stderr.writeln('WARNING: failed to parse "text" property out of build JSON');
-    return false;
-  }
-
-  return buildJson['text'][1] == 'successful';
-}
diff --git a/agent/lib/src/commands/ci.dart b/agent/lib/src/commands/ci.dart
index 2554a6de94..27652ee84e 100644
--- a/agent/lib/src/commands/ci.dart
+++ b/agent/lib/src/commands/ci.dart
@@ -3,7 +3,6 @@
 // found in the LICENSE file.
 
 import 'dart:async';
-import 'dart:convert';
 import 'dart:io';
 
 import 'package:args/args.dart';
@@ -11,7 +10,8 @@ import 'package:args/args.dart';
 import '../adb.dart';
 import '../agent.dart';
 import '../firebase.dart';
-import '../framework.dart';
+import '../golem.dart';
+import '../runner.dart';
 import '../utils.dart';
 
 /// Agents periodically poll the server for more tasks. This sleep period is
@@ -31,6 +31,10 @@ class ContinuousIntegrationCommand extends Command {
 
   @override
   Future<Null> run(ArgResults args) async {
+    return runAndCaptureAsyncStacks(() => _runContinuously(args));
+  }
+
+  Future<Null> _runContinuously(ArgResults args) async {
     // Perform one pre-flight round of checks and quit immediately if something
     // is wrong.
     AgentHealth health = await _performHealthChecks();
@@ -66,24 +70,15 @@ class ContinuousIntegrationCommand extends Command {
         CocoonTask task = await agent.reserveTask();
         try {
           if (task != null) {
+            section('Task info:');
+            print('  name     : ${task.name}');
+            print('  key      : ${task.key ?? ""}');
+            print('  revision : ${task.revision}');
+
             // Sync flutter outside of the task so it does not contribute to
             // the task timeout.
             await getFlutterAt(task.revision).timeout(const Duration(minutes: 10));
-
-            // No need to pass revision as repo syncing is done here.
-            List<String> runnerArgs = <String>[
-              'run',
-              '--task-name=${task.name}',
-              '--task-key=${task.key}',
-            ];
-
-            Process proc = await startProcess(
-              dartBin,
-              [config.runTaskFile.path]..addAll(runnerArgs)
-            ).timeout(const Duration(minutes: 1));
-
-            _logStandardStreams(task, proc);
-            await proc.exitCode.timeout(taskTimeoutWithGracePeriod);
+            await _runTask(task);
           }
         } catch(error, stackTrace) {
           print('ERROR: $error\n$stackTrace');
@@ -92,6 +87,7 @@ class ContinuousIntegrationCommand extends Command {
       } catch(error, stackTrace) {
         print('ERROR: $error\n$stackTrace');
       } finally {
+        await _screensOff();
         await forceQuitRunningProcesses();
       }
 
@@ -99,6 +95,29 @@ class ContinuousIntegrationCommand extends Command {
     }
   }
 
+  Future<Null> _runTask(CocoonTask task) async {
+    await runAndCaptureAsyncStacks(() async {
+      TaskResult result = await runTask(agent, task);
+      if (result.succeeded) {
+        await agent.updateTaskStatus(task.key, 'Succeeded');
+        await _uploadDataToFirebase(task, result);
+      } else {
+        await agent.updateTaskStatus(task.key, 'Failed');
+      }
+    });
+  }
+
+  Future<Null> _screensOff() async {
+    try {
+      for (Device device in await devices.discoverDevices()) {
+        device.sendToSleep();
+      }
+    } catch(error, stackTrace) {
+      // Best effort only.
+      print('Failed to turn off screen: $error\n$stackTrace');
+    }
+  }
+
   Future<AgentHealth> _performHealthChecks() async {
     AgentHealth results = new AgentHealth();
     try {
@@ -138,31 +157,6 @@ class ContinuousIntegrationCommand extends Command {
     return results;
   }
 
-  /// Listens to standard output and upload logs to Cocoon in semi-realtime.
-  Future<Null> _logStandardStreams(CocoonTask task, Process proc) async {
-    StringBuffer buffer = new StringBuffer();
-
-    Future<Null> sendLog(String message, {bool flush: false}) async {
-      buffer.write(message);
-      print('[task runner] $message');
-      // Send in chunks 100KB each, or upon request.
-      if (flush || buffer.length > 100000) {
-        String chunk = buffer.toString();
-        buffer = new StringBuffer();
-        await agent.uploadLogChunk(task, chunk);
-      }
-    }
-
-    proc.stdout.transform(UTF8.decoder).listen((String s) {
-      sendLog(s);
-    });
-    proc.stderr.transform(UTF8.decoder).listen((String s) {
-      sendLog(s);
-    });
-    await proc.exitCode;
-    sendLog('Task execution finished', flush: true);
-  }
-
   void _listenToShutdownSignals() {
     _streamSubscriptions.addAll(<StreamSubscription>[
       ProcessSignal.SIGINT.watch().listen((_) {
@@ -220,3 +214,33 @@ Future<HealthCheckResult> _scrapeRemoteAccessInfo() async {
     : 'Possible remote access IP: $ip'
   );
 }
+
+Future<Null> _uploadDataToFirebase(CocoonTask task, TaskResult result) async {
+  List<Map<String, dynamic>> golemData = <Map<String, dynamic>>[];
+  int golemRevision = await computeGolemRevision();
+
+  Map<String, dynamic> data = new Map<String, dynamic>.from(result.data);
+
+  if (result.benchmarkScoreKeys != null) {
+    for (String scoreKey in result.benchmarkScoreKeys) {
+      String benchmarkName = '${task.name}.$scoreKey';
+      if (registeredBenchmarkNames.contains(benchmarkName)) {
+        golemData.add(<String, dynamic>{
+          'benchmark_name': benchmarkName,
+          'golem_revision': golemRevision,
+          'score': result.data[scoreKey],
+        });
+      }
+    }
+  }
+
+  data['__metadata__'] = <String, dynamic>{
+    'success': result.succeeded,
+    'revision': task.revision,
+    'message': result.reason ?? 'N/A',
+  };
+
+  data['__golem__'] = golemData;
+
+  await uploadToFirebase(task.name, data);
+}
diff --git a/agent/lib/src/commands/run.dart b/agent/lib/src/commands/run.dart
deleted file mode 100644
index 6192a5175c..0000000000
--- a/agent/lib/src/commands/run.dart
+++ /dev/null
@@ -1,124 +0,0 @@
-// Copyright 2016 The Chromium Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style license that can be
-// found in the LICENSE file.
-
-import 'dart:async';
-import 'dart:io';
-
-import 'package:args/args.dart';
-import 'package:stack_trace/stack_trace.dart';
-
-import '../adb.dart';
-import '../agent.dart';
-import '../golem.dart';
-import '../firebase.dart';
-import '../framework.dart';
-import '../utils.dart';
-
-/// Runs the agent once to perform a specific task and, optionally, report the
-/// result back to Cocoon backend.
-///
-/// The task is run at the currently synced revision of Flutter. This command
-/// assumes that Flutter is already synced to the desired revision.
-class RunCommand extends Command {
-  RunCommand(Agent agent) : super('run', agent);
-
-  static final ArgParser argParser = new ArgParser()
-    ..addOption(
-      'task-name',
-      help: '(required) The name of the task to run.'
-    )
-    ..addOption(
-      'task-key',
-      help: '(optional) The key of the task to update the status of in Cocoon. '
-            'It is only required if you want the runner to upload logs and '
-            'change task status in Cocoon.'
-    );
-
-  @override
-  Future<Null> run(ArgResults args) async {
-    CocoonTask task = new CocoonTask(
-      name: args['task-name'],
-      key: args['task-key'],
-      revision: await getCurrentFlutterRepoCommit()
-    );
-
-    section('Task info:');
-    print('  name     : ${task.name}');
-    print('  key      : ${task.key ?? ""}');
-    print('  revision : ${task.revision}');
-
-    if (task.name == null || task.name == '') {
-      print('\n Incorrect command-line options. Usage:');
-      print(argParser.usage);
-      exit(1);
-    }
-
-    try {
-      await runAndCaptureAsyncStacks(() async {
-        // Load-balance tests across attached devices
-        await devices.chooseWorkingDevice();
-
-        BuildResult result = await agent.performTask(task);
-        if (task.key != null) {
-          if (result.succeeded) {
-            await agent.updateTaskStatus(task.key, 'Succeeded');
-            await _uploadDataToFirebase(result);
-          } else {
-            await agent.updateTaskStatus(task.key, 'Failed');
-          }
-        }
-      }).timeout(taskTimeout);
-    } catch(error, stackTrace) {
-      print('Caught: $error\n${new Chain.forTrace(stackTrace).terse}');
-      if (task.key != null)
-        await agent.updateTaskStatus(task.key, 'Failed');
-      exitCode = 1;
-    } finally {
-      await forceQuitRunningProcesses();
-    }
-
-    if (exitCode != 0) {
-      // Force-quitting child processes sometimes causes the Dart VM to fail to
-      // exit. So we give a 2-second grace period for any pending cleanups, then
-      // self-destruct.
-      await new Future<Null>.delayed(const Duration(seconds: 2));
-      exit(exitCode);
-    }
-  }
-}
-
-Future<Null> _uploadDataToFirebase(BuildResult result) async {
-  List<Map<String, dynamic>> golemData = <Map<String, dynamic>>[];
-
-  for (TaskResult taskResult in result.results) {
-    // TODO(devoncarew): We should also upload the fact that these tasks failed.
-    if (taskResult.data == null)
-      continue;
-
-    Map<String, dynamic> data = new Map<String, dynamic>.from(taskResult.data.json);
-
-    if (taskResult.data.benchmarkScoreKeys != null) {
-      for (String scoreKey in taskResult.data.benchmarkScoreKeys) {
-        String benchmarkName = '${taskResult.task.name}.$scoreKey';
-        if (registeredBenchmarkNames.contains(benchmarkName)) {
-          golemData.add(<String, dynamic>{
-            'benchmark_name': benchmarkName,
-            'golem_revision': result.golemRevision,
-            'score': taskResult.data.json[scoreKey],
-          });
-        }
-      }
-    }
-
-    data['__metadata__'] = <String, dynamic>{
-      'success': taskResult.succeeded,
-      'revision': taskResult.revision,
-      'message': taskResult.message,
-    };
-
-    data['__golem__'] = golemData;
-
-    uploadToFirebase(taskResult.task.name, data);
-  }
-}
diff --git a/agent/lib/src/framework.dart b/agent/lib/src/framework.dart
deleted file mode 100644
index ff18ba235d..0000000000
--- a/agent/lib/src/framework.dart
+++ /dev/null
@@ -1,161 +0,0 @@
-// Copyright 2016 The Chromium Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style license that can be
-// found in the LICENSE file.
-
-import 'dart:async';
-import 'dart:io';
-import 'dart:convert';
-
-import 'utils.dart';
-
-/// Maximum amount of time a single task is allowed to take to run.
-///
-/// If exceeded the task is considered to have failed.
-const Duration taskTimeout = const Duration(minutes: 10);
-
-/// Slightly longer than [taskTimeout] that gives the task runner a chance to
-/// clean-up before forcefully quitting it.
-const Duration taskTimeoutWithGracePeriod = const Duration(minutes: 11);
-
-/// Represents a unit of work performed on the dashboard build box that can
-/// succeed, fail and be retried independently of others.
-abstract class Task {
-  Task(this.name);
-
-  /// The name of the task that shows up in log messages.
-  ///
-  /// This should be as unique as possible to avoid confusion.
-  final String name;
-
-  /// Performs actual work.
-  Future<TaskResultData> run();
-}
-
-/// Runs a queue of tasks; collects results.
-class TaskRunner {
-  TaskRunner(this.revision, this.golemRevision, this.tasks);
-
-  /// Flutter repository revision at which this runner ran.
-  final String revision;
-
-  /// Revision _number_ as understood by Golem.
-  ///
-  /// See [computeGolemRevision].
-  final int golemRevision;
-  final List<Task> tasks;
-
-  Future<BuildResult> run() async {
-    List<TaskResult> results = <TaskResult>[];
-
-    for (Task task in tasks) {
-      section('Running task ${task.name}');
-      TaskResult result;
-      try {
-        TaskResultData data = await task.run();
-        if (data != null)
-          result = new TaskResult.success(task, revision, data);
-        else
-          result = new TaskResult.failure(task, revision, 'Task data missing');
-      } catch (taskError, taskErrorStack) {
-        String message = '${task.name} failed: $taskError';
-        if (taskErrorStack != null) {
-          message += '\n\n$taskErrorStack';
-        }
-        print('');
-        print(message);
-        result = new TaskResult.failure(task, revision, message);
-      }
-      results.add(result);
-      section('Task ${task.name} ${result.succeeded ? "succeeded" : "failed"}.');
-    }
-
-    return new BuildResult(golemRevision, results);
-  }
-}
-
-/// All results accumulated from a build session.
-class BuildResult {
-  BuildResult(this.golemRevision, this.results);
-
-  final int golemRevision;
-
-  /// Individual task results.
-  final List<TaskResult> results;
-
-  /// Whether the overall build failed.
-  ///
-  /// We consider the build as failed if at least one task fails.
-  bool get failed => results.any((TaskResult res) => res.failed);
-
-  /// The opposite of [failed], i.e. all tasks succeeded.
-  bool get succeeded => !failed;
-
-  /// The number of failed tasks.
-  int get failedTaskCount => results.fold(0, (int previous, TaskResult res) => previous + (res.failed ? 1 : 0));
-}
-
-/// A result of running a single task.
-class TaskResult {
-
-  /// Constructs a successful result.
-  TaskResult.success(this.task, this.revision, this.data)
-    : this.succeeded = true,
-      this.message = 'success';
-
-  /// Constructs an unsuccessful result.
-  TaskResult.failure(this.task, this.revision, this.message)
-    : this.succeeded = false,
-      this.data = null;
-
-  /// The task that was run.
-  final Task task;
-
-  /// Whether the task succeeded.
-  final bool succeeded;
-
-  /// The revision of Flutter repo at which this result was generated.
-  final String revision;
-
-  /// Data generated by the task that will be uploaded to Firebase.
-  final TaskResultData data;
-
-  /// Whether the task failed.
-  bool get failed => !succeeded;
-
-  /// Explains the result in a human-readable format.
-  final String message;
-}
-
-/// Data generated by the task that will be uploaded to Firebase.
-class TaskResultData {
-  TaskResultData(this.json, {this.benchmarkScoreKeys}) {
-    const JsonEncoder prettyJson = const JsonEncoder.withIndent('  ');
-    if (benchmarkScoreKeys != null) {
-      for (String key in benchmarkScoreKeys) {
-        if (!json.containsKey(key)) {
-          throw 'Invalid Golem score key "$key". It does not exist in task result data ${prettyJson.convert(json)}';
-        }
-      }
-    }
-  }
-
-  factory TaskResultData.fromFile(File file, {List<String> benchmarkScoreKeys}) {
-    return new TaskResultData(JSON.decode(file.readAsStringSync()), benchmarkScoreKeys: benchmarkScoreKeys);
-  }
-
-  /// Task-specific JSON data
-  final Map<String, dynamic> json;
-
-  /// Keys in [json] that store scores that will be submitted to Golem.
-  ///
-  /// Each key is also part of a benchmark's name tracked by Golem.
-  /// A benchmark name is computed by combining [Task.name] with a key
-  /// separated by a dot. For example, if a task's name is
-  /// `"complex_layout__start_up"` and score key is
-  /// `"engineEnterTimestampMicros"`, the score will be submitted to Golem under
-  /// `"complex_layout__start_up.engineEnterTimestampMicros"`.
-  ///
-  /// This convention reduces the amount of configuration that needs to be done
-  /// to submit benchmark scores to Golem.
-  final List<String> benchmarkScoreKeys;
-}
diff --git a/agent/lib/src/gallery.dart b/agent/lib/src/gallery.dart
deleted file mode 100644
index a03c594ff6..0000000000
--- a/agent/lib/src/gallery.dart
+++ /dev/null
@@ -1,55 +0,0 @@
-// Copyright 2016 The Chromium Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style license that can be
-// found in the LICENSE file.
-
-import 'dart:async';
-import 'dart:convert';
-import 'dart:io';
-
-import 'adb.dart';
-import 'framework.dart';
-import 'utils.dart';
-
-Task createGalleryTransitionTest({ bool ios: false }) => new GalleryTransitionTest(ios: ios);
-
-class GalleryTransitionTest extends Task {
-  GalleryTransitionTest({ bool ios }) :
-    this.ios = ios,
-    super('flutter_gallery${ios ? "_ios" : ""}__transition_perf');
-
-  final bool ios;
-
-  @override
-  Future<TaskResultData> run() async {
-    String deviceId = devices.workingDevice.deviceId;
-    Directory galleryDirectory = dir('${config.flutterDirectory.path}/examples/flutter_gallery');
-    await inDirectory(galleryDirectory, () async {
-      await flutter('packages', options: ['get']);
-
-      if (ios) {
-        // This causes an Xcode project to be created.
-        await flutter('build', options: ['ios', '--profile']);
-      }
-
-      await flutter('drive', options: [
-        '--profile',
-        '--trace-startup',
-        '-t',
-        'test_driver/transitions_perf.dart',
-        '-d',
-        deviceId,
-      ]);
-    });
-
-    // Route paths contains slashes, which Firebase doesn't accept in keys, so we
-    // remove them.
-    Map<String, dynamic> original = JSON.decode(file('${galleryDirectory.path}/build/transition_durations.timeline.json').readAsStringSync());
-    Map<String, dynamic> clean = new Map.fromIterable(
-      original.keys,
-      key: (String key) => key.replaceAll('/', ''),
-      value: (String key) => original[key]
-    );
-
-    return new TaskResultData(clean);
-  }
-}
diff --git a/agent/lib/src/hot_dev_cycle.dart b/agent/lib/src/hot_dev_cycle.dart
deleted file mode 100644
index b0208afd2d..0000000000
--- a/agent/lib/src/hot_dev_cycle.dart
+++ /dev/null
@@ -1,56 +0,0 @@
-// Copyright (c) 2016 The Chromium Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style license that can be
-// found in the LICENSE file.
-
-import 'dart:async';
-import 'dart:io';
-
-import 'package:path/path.dart' as path;
-
-import 'adb.dart';
-import 'framework.dart';
-import 'utils.dart';
-
-Task createHotDevCycleTest({
-  String commit,
-  DateTime timestamp
-}) => new HotDevCycleTask(commit, timestamp);
-
-class HotDevCycleTask extends Task {
-  HotDevCycleTask(this.commit, this.timestamp)
-      : super('hot_mode_dev_cycle__benchmark') {
-    assert(commit != null);
-    assert(timestamp != null);
-  }
-
-  final String commit;
-  final DateTime timestamp;
-
-  Directory get appDir =>
-      dir(path.join(config.flutterDirectory.path,
-          'examples/flutter_gallery'));
-
-  File get benchmarkFile =>
-      file(path.join(appDir.path, 'hot_benchmark.json'));
-
-  final List<String> benchmarkScoreKeys = [
-    'hotReloadMillisecondsToFrame',
-    'hotRestartMillisecondsToFrame'
-  ];
-
-  @override
-  Future<TaskResultData> run() async {
-    Device device = devices.workingDevice;
-    device.unlock();
-    rm(benchmarkFile);
-    await inDirectory(appDir, () async {
-      return await flutter(
-        'run',
-        options: ['--hot', '-d', device.deviceId, '--benchmark'],
-        canFail: false
-      );
-    });
-    return new TaskResultData.fromFile(benchmarkFile,
-                                       benchmarkScoreKeys: benchmarkScoreKeys);
-  }
-}
diff --git a/agent/lib/src/perf_tests.dart b/agent/lib/src/perf_tests.dart
deleted file mode 100644
index d2e86f2098..0000000000
--- a/agent/lib/src/perf_tests.dart
+++ /dev/null
@@ -1,167 +0,0 @@
-// Copyright (c) 2016 The Chromium Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style license that can be
-// found in the LICENSE file.
-
-import 'dart:async';
-import 'dart:convert' show JSON;
-
-import 'adb.dart';
-import 'framework.dart';
-import 'utils.dart';
-
-Task createComplexLayoutScrollPerfTest({ bool ios: false }) {
-  return new PerfTest(
-    'complex_layout_scroll_perf${ios ? "_ios" : ""}__timeline_summary',
-    '${config.flutterDirectory.path}/dev/benchmarks/complex_layout',
-    'test_driver/scroll_perf.dart',
-    'complex_layout_scroll_perf',
-    ios: ios
-  );
-}
-
-Task createFlutterGalleryStartupTest({ bool ios: false }) {
-  return new StartupTest(
-    'flutter_gallery${ios ? "_ios" : ""}__start_up',
-    '${config.flutterDirectory.path}/examples/flutter_gallery',
-    ios: ios
-  );
-}
-
-Task createComplexLayoutStartupTest({ bool ios: false }) {
-  return new StartupTest(
-    'complex_layout${ios ? "_ios" : ""}__start_up',
-    '${config.flutterDirectory.path}/dev/benchmarks/complex_layout',
-    ios: ios
-  );
-}
-
-Task createFlutterGalleryBuildTest() {
-  return new BuildTest('flutter_gallery__build', '${config.flutterDirectory.path}/examples/flutter_gallery');
-}
-
-Task createComplexLayoutBuildTest() {
-  return new BuildTest('complex_layout__build', '${config.flutterDirectory.path}/dev/benchmarks/complex_layout');
-}
-
-/// Measure application startup performance.
-class StartupTest extends Task {
-  static const Duration _startupTimeout = const Duration(minutes: 2);
-
-  StartupTest(String name, this.testDirectory, { this.ios }) : super(name);
-
-  final String testDirectory;
-  final bool ios;
-
-  Future<TaskResultData> run() async {
-    return await inDirectory(testDirectory, () async {
-      String deviceId = devices.workingDevice.deviceId;
-      await flutter('packages', options: ['get']);
-
-      if (ios) {
-        // This causes an Xcode project to be created.
-        await flutter('build', options: ['ios', '--profile']);
-      }
-
-      await flutter('run', options: [
-        '--profile',
-        '--trace-startup',
-        '-d',
-        deviceId
-      ]).timeout(_startupTimeout);
-      Map<String, dynamic> data = JSON.decode(file('$testDirectory/build/start_up_info.json').readAsStringSync());
-      return new TaskResultData(data, benchmarkScoreKeys: <String>[
-        'engineEnterTimestampMicros',
-        'timeToFirstFrameMicros',
-      ]);
-    });
-  }
-}
-
-/// Measures application runtime performance, specifically per-frame
-/// performance.
-class PerfTest extends Task {
-
-  PerfTest(String name, this.testDirectory, this.testTarget, this.timelineFileName, { this.ios })
-      : super(name);
-
-  final String testDirectory;
-  final String testTarget;
-  final String timelineFileName;
-  final bool ios;
-
-  @override
-  Future<TaskResultData> run() {
-    return inDirectory(testDirectory, () async {
-      String deviceId = devices.workingDevice.deviceId;
-      await flutter('packages', options: ['get']);
-
-      if (ios) {
-        // This causes an Xcode project to be created.
-        await flutter('build', options: ['ios', '--profile']);
-      }
-
-      await flutter('drive', options: [
-        '-v',
-        '--profile',
-        '--trace-startup', // Enables "endless" timeline event buffering.
-        '-t',
-        testTarget,
-        '-d',
-        deviceId,
-      ]);
-      Map<String, dynamic> data = JSON.decode(file('$testDirectory/build/${timelineFileName}.timeline_summary.json').readAsStringSync());
-      return new TaskResultData(data, benchmarkScoreKeys: <String>[
-        'average_frame_build_time_millis',
-        'worst_frame_build_time_millis',
-        'missed_frame_build_budget_count',
-      ]);
-    });
-  }
-}
-
-class BuildTest extends Task {
-
-  BuildTest(String name, this.testDirectory) : super(name);
-
-  final String testDirectory;
-
-  Future<TaskResultData> run() async {
-    return await inDirectory(testDirectory, () async {
-      Device device = devices.workingDevice;
-      device.unlock();
-      await flutter('packages', options: ['get']);
-
-      var watch = new Stopwatch()..start();
-      await flutter('build', options: [
-        'aot',
-        '--profile',
-        '--no-pub',
-        '--target-platform', 'android-arm'  // Generate blobs instead of assembly.
-      ]);
-      watch.stop();
-
-      var vmisolateSize = file("$testDirectory/build/aot/snapshot_aot_vmisolate").lengthSync();
-      var isolateSize = file("$testDirectory/build/aot/snapshot_aot_isolate").lengthSync();
-      var instructionsSize = file("$testDirectory/build/aot/snapshot_aot_instr").lengthSync();
-      var rodataSize = file("$testDirectory/build/aot/snapshot_aot_rodata").lengthSync();
-      var totalSize = vmisolateSize + isolateSize + instructionsSize + rodataSize;
-
-      Map<String, dynamic> data = {
-        'aot_snapshot_build_millis': watch.elapsedMilliseconds,
-        'aot_snapshot_size_vmisolate': vmisolateSize,
-        'aot_snapshot_size_isolate': isolateSize,
-        'aot_snapshot_size_instructions': instructionsSize,
-        'aot_snapshot_size_rodata': rodataSize,
-        'aot_snapshot_size_total': totalSize,
-      };
-      return new TaskResultData(data, benchmarkScoreKeys: <String>[
-        'aot_snapshot_build_millis',
-        'aot_snapshot_size_vmisolate',
-        'aot_snapshot_size_isolate',
-        'aot_snapshot_size_instructions',
-        'aot_snapshot_size_rodata',
-        'aot_snapshot_size_total',
-      ]);
-    });
-  }
-}
diff --git a/agent/lib/src/refresh.dart b/agent/lib/src/refresh.dart
deleted file mode 100644
index a34592d601..0000000000
--- a/agent/lib/src/refresh.dart
+++ /dev/null
@@ -1,73 +0,0 @@
-// Copyright (c) 2016 The Chromium Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style license that can be
-// found in the LICENSE file.
-
-import 'dart:async';
-import 'dart:io';
-
-import 'package:path/path.dart' as path;
-
-import 'adb.dart';
-import 'benchmarks.dart';
-import 'framework.dart';
-import 'utils.dart';
-
-Task createRefreshTest({
-  String commit,
-  DateTime timestamp
-}) => new EditRefreshTask(commit, timestamp);
-
-class EditRefreshTask extends Task {
-  EditRefreshTask(this.commit, this.timestamp)
-      : super('mega_gallery__refresh_time') {
-    assert(commit != null);
-    assert(timestamp != null);
-  }
-
-  final String commit;
-  final DateTime timestamp;
-
-  @override
-  Future<TaskResultData> run() async {
-    Device device = devices.workingDevice;
-    device.unlock();
-    Benchmark benchmark = new EditRefreshBenchmark(commit, timestamp);
-    section(benchmark.name);
-    await runBenchmark(benchmark, iterations: 3, warmUpBenchmark: true);
-    return benchmark.bestResult;
-  }
-}
-
-class EditRefreshBenchmark extends Benchmark {
-  EditRefreshBenchmark(this.commit, this.timestamp)
-      : super('edit refresh');
-
-  final String commit;
-  final DateTime timestamp;
-
-  Directory get megaDir => dir(path.join(config.flutterDirectory.path, 'dev/benchmarks/mega_gallery'));
-  File get benchmarkFile => file(path.join(megaDir.path, 'refresh_benchmark.json'));
-
-  @override
-  TaskResultData get lastResult => new TaskResultData.fromFile(benchmarkFile);
-
-  Future<Null> init() {
-    return inDirectory(config.flutterDirectory, () async {
-      await dart(['dev/tools/mega_gallery.dart']);
-    });
-  }
-
-  @override
-  Future<num> run() async {
-    Device device = devices.workingDevice;
-    rm(benchmarkFile);
-    int exitCode = await inDirectory(megaDir, () async {
-      return await flutter(
-        'run', options: ['-d', device.deviceId, '--benchmark'], canFail: true
-      );
-    });
-    if (exitCode != 0)
-      return new Future.error(exitCode);
-    return addBuildInfo(benchmarkFile, timestamp: timestamp, expected: 200, commit: commit);
-  }
-}
diff --git a/agent/lib/src/runner.dart b/agent/lib/src/runner.dart
new file mode 100644
index 0000000000..bf7e9dca0e
--- /dev/null
+++ b/agent/lib/src/runner.dart
@@ -0,0 +1,188 @@
+// Copyright 2016 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+import 'dart:async';
+import 'dart:convert';
+import 'dart:io';
+
+import 'package:vm_service_client/vm_service_client.dart';
+
+import 'agent.dart';
+import 'utils.dart';
+
+/// Slightly longer task timeout that gives the task runner a chance to
+/// clean-up before forcefully quitting it.
+const Duration taskTimeoutWithGracePeriod = const Duration(minutes: 11);
+
+/// Send logs in 10KB chunks.
+const int _kLogChunkSize = 10000;
+
+/// A result of running a single task.
+///
+/// In normal circumstances, even when a task fails, the result is parsed from
+/// JSON returned by the task runner process via the VM service. However, if
+/// things are completely out of control and the task runner process is
+/// corrupted a failed result can be instantiated directly using
+/// [TaskResult.failure] constructor.
+class TaskResult {
+  /// Parses a task result from JSON.
+  TaskResult.parse(Map<String, dynamic> json)
+    : succeeded = json['success'],
+      data = json['data'],
+      benchmarkScoreKeys = json['benchmarkScoreKeys'] ?? const <String>[],
+      reason = json['reason'];
+
+  /// Constructs an unsuccessful result.
+  TaskResult.failure(this.reason)
+      : this.succeeded = false,
+        this.data = const <String, dynamic>{},
+        this.benchmarkScoreKeys = const <String>[];
+
+  /// Whether the task succeeded.
+  final bool succeeded;
+
+  /// Task-specific JSON data.
+  final Map<String, dynamic> data;
+
+  /// Keys in [data] that store scores that will be submitted to Golem.
+  ///
+  /// Each key is also part of a benchmark's name tracked by Golem.
+  /// A benchmark name is computed by combining [Task.name] with a key
+  /// separated by a dot. For example, if a task's name is
+  /// `"complex_layout__start_up"` and score key is
+  /// `"engineEnterTimestampMicros"`, the score will be submitted to Golem under
+  /// `"complex_layout__start_up.engineEnterTimestampMicros"`.
+  ///
+  /// This convention reduces the amount of configuration that needs to be done
+  /// to submit benchmark scores to Golem.
+  final List<String> benchmarkScoreKeys;
+
+  /// Whether the task failed.
+  bool get failed => !succeeded;
+
+  /// Explains the failure reason if [failed].
+  final String reason;
+}
+
+/// Runs a task in a separate Dart VM and collects the result using the VM
+/// service protocol.
+///
+/// [taskName] is the name of the task. The corresponding task executable is
+/// expected to be found under `bin/tasks`.
+Future<TaskResult> runTask(Agent agent, CocoonTask task) async {
+  String devicelabPath = '${config.flutterDirectory.path}/dev/devicelab';
+  String taskExecutable = 'bin/tasks/${task.name}.dart';
+
+  if (!file('$devicelabPath/$taskExecutable').existsSync())
+    throw 'Executable Dart file not found: $taskExecutable';
+
+  int vmServicePort = await _findAvailablePort();
+  Process runner;
+  await inDirectory(devicelabPath, () async {
+    runner = await startProcess(dartBin, <String>[
+      '--enable-vm-service=$vmServicePort',
+      '--no-pause-isolates-on-exit',
+      taskExecutable,
+    ]);
+  });
+
+  bool runnerFinished = false;
+
+  runner.exitCode.then((_) {
+    runnerFinished = true;
+  });
+
+  StringBuffer buffer = new StringBuffer();
+
+  Future<Null> sendLog(String message, {bool flush: false}) async {
+    buffer.write(message);
+    print('[task runner] [${task.name}] $message');
+    // Send a chunk at a time, or upon request.
+    if (flush || buffer.length > _kLogChunkSize) {
+      String chunk = buffer.toString();
+      buffer = new StringBuffer();
+      await agent.uploadLogChunk(task, chunk);
+    }
+  }
+
+  var stdoutSub = runner.stdout.transform(UTF8.decoder).listen(sendLog);
+  var stderrSub = runner.stderr.transform(UTF8.decoder).listen(sendLog);
+
+  String waitingFor = 'connection';
+  try {
+    VMIsolate isolate = await _connectToRunnerIsolate(vmServicePort);
+    waitingFor = 'task completion';
+    Map<String, dynamic> taskResult =
+        await isolate.invokeExtension('ext.cocoonRunTask').timeout(taskTimeoutWithGracePeriod);
+    waitingFor = 'task process to exit';
+    await runner.exitCode.timeout(const Duration(seconds: 1));
+    return new TaskResult.parse(taskResult);
+  } on TimeoutException catch (timeout) {
+    runner.kill(ProcessSignal.SIGINT);
+    return new TaskResult.failure(
+      'Timeout waiting for $waitingFor: ${timeout.message}'
+    );
+  } finally {
+    await stdoutSub.cancel();
+    await stderrSub.cancel();
+    sendLog('Task execution finished', flush: true);
+    if (!runnerFinished)
+      runner.kill(ProcessSignal.SIGKILL);
+    await forceQuitRunningProcesses();
+  }
+}
+
+Future<VMIsolate> _connectToRunnerIsolate(int vmServicePort) async {
+  String url = 'ws://localhost:$vmServicePort/ws';
+  DateTime started = new DateTime.now();
+
+  // TODO(yjbanov): due to lack of imagination at the moment the handshake with
+  //                the task process is very rudimentary and requires this small
+  //                delay to let the task process open up the VM service port.
+  //                Otherwise we almost always hit the non-ready case first and
+  //                wait a whole 1 second, which is annoying.
+  await new Future<Null>.delayed(const Duration(milliseconds: 100));
+
+  while (true) {
+    try {
+      // Make sure VM server is up by successfully opening and closing a socket.
+      await (await WebSocket.connect(url)).close();
+
+      // Look up the isolate.
+      VMServiceClient client = new VMServiceClient.connect(url);
+      VM vm = await client.getVM();
+      VMIsolate isolate = vm.isolates.single;
+      String response = await isolate.invokeExtension('ext.cocoonRunnerReady');
+      if (response != 'ready')
+        throw 'not ready yet';
+      return isolate;
+    } catch (error) {
+      const Duration connectionTimeout = const Duration(seconds: 2);
+      if (new DateTime.now().difference(started) > connectionTimeout) {
+        throw new TimeoutException(
+          'Failed to connect to the task runner process',
+          connectionTimeout,
+        );
+      }
+      print('VM service not ready yet: $error');
+      const Duration pauseBetweenRetries = const Duration(milliseconds: 200);
+      print('Will retry in $pauseBetweenRetries.');
+      await new Future<Null>.delayed(pauseBetweenRetries);
+    }
+  }
+}
+
+Future<int> _findAvailablePort() async {
+  int port = 20000;
+  while (true) {
+    try {
+      ServerSocket socket =
+          await ServerSocket.bind(InternetAddress.LOOPBACK_IP_V4, port);
+      await socket.close();
+      return port;
+    } catch (_) {
+      port++;
+    }
+  }
+}
diff --git a/agent/lib/src/size_tests.dart b/agent/lib/src/size_tests.dart
deleted file mode 100644
index cb911cf7bb..0000000000
--- a/agent/lib/src/size_tests.dart
+++ /dev/null
@@ -1,46 +0,0 @@
-// Copyright 2016 The Chromium Authors. All rights reserved.
-// Use of this source code is governed by a BSD-style license that can be
-// found in the LICENSE file.
-
-import 'dart:async';
-import 'dart:io';
-
-import 'framework.dart';
-import 'utils.dart';
-
-Task createBasicMaterialAppSizeTest() => new BasicMaterialAppSizeTest();
-
-class BasicMaterialAppSizeTest extends Task {
-  BasicMaterialAppSizeTest() : super('basic_material_app__size');
-
-  @override
-  Future<TaskResultData> run() async {
-    const sampleAppName = 'sample_flutter_app';
-    Directory sampleDir = dir('${Directory.systemTemp.path}/$sampleAppName');
-
-    if (await sampleDir.exists())
-      rrm(sampleDir);
-
-    int apkSizeInBytes;
-
-    await inDirectory(Directory.systemTemp, () async {
-      await flutter('create', options: [sampleAppName]);
-
-      if (!(await sampleDir.exists()))
-        throw 'Failed to create sample Flutter app in ${sampleDir.path}';
-
-      await inDirectory(sampleDir, () async {
-        await flutter('packages', options: ['get']);
-        await flutter('build', options: ['clean']);
-        await flutter('build', options: ['apk', '--release']);
-        apkSizeInBytes = await file('${sampleDir.path}/build/app.apk').length();
-      });
-    });
-
-    return new TaskResultData({
-      'release_size_in_bytes': apkSizeInBytes
-    }, benchmarkScoreKeys: [
-      'release_size_in_bytes'
-    ]);
-  }
-}
diff --git a/agent/pubspec.yaml b/agent/pubspec.yaml
index e07d96e722..00136cfff7 100644
--- a/agent/pubspec.yaml
+++ b/agent/pubspec.yaml
@@ -1,7 +1,7 @@
 name: cocoon_agent
 version: 0.0.1
 author: Flutter Authors <flutter-dev@googlegroups.com>
-description: Scripts that run Cocoon checklist tasks
+description: Flutter continuous integration agent (Cocoon agent)
 homepage: https://github.com/flutter/cocoon
 
 environment:
@@ -9,21 +9,14 @@ environment:
 
 dependencies:
   args: ^0.13.0
-  browser: ^0.10.0
-  charted: ^0.4.0
-  dart_to_js_script_rewriter: ^1.0.0
-  firebase: ^0.6.6
   firebase_rest: ^0.1.3
   http: ^0.11.3
-  intl: ^0.12.7
   meta: ^0.12.1
   path: ^1.3.0
   stack_trace: ^1.6.5
+  vm_service_client: ^0.2.0
   yaml: ^2.1.10
 
 dev_dependencies:
   test: ^0.12.13
   collection: ^1.8.0
-
-transformers:
-  - dart_to_js_script_rewriter
diff --git a/app/bin/build_and_test.sh b/app/bin/build_and_test.sh
index da1af910ed..a5afb6a3ee 100755
--- a/app/bin/build_and_test.sh
+++ b/app/bin/build_and_test.sh
@@ -10,7 +10,7 @@ if [ ! -f "pubspec.yaml" -a ! -f "app.yaml" ]; then
 fi
 
 rm -rf build
-pub get --no-packages-dir
+pub get
 pub run test
 pub build
 cp web/*.dart build/web/
diff --git a/db/schema.go b/db/schema.go
index 7b477b4c85..4fb38c2724 100644
--- a/db/schema.go
+++ b/db/schema.go
@@ -115,9 +115,10 @@ type Agent struct {
 	AgentID              string
 	IsHealthy            bool
 	HealthCheckTimestamp int64
-	HealthDetails        string // a human-readable printout of health details
-	AuthTokenHash        []byte
-	Capabilities         []string
+	// a human-readable printout of health details
+	HealthDetails string `datastore:"HealthDetails,noindex"`
+	AuthTokenHash []byte
+	Capabilities  []string
 }
 
 // AgentStatus contains agent health status.