Agent Patterns Demo Pack — Agent Reference

This document describes every agent across all seven demos, the orchestration pattern each demo uses, and how to configure the model they run on.

Model Configuration

All agents are powered by a single, runtime-switchable model provider. Choose between Foundry Local (on-device, no API key) or Microsoft Foundry (cloud). Switch providers and models live from the ⚙ Model Settings panel in the launcher UI without restarting the app.

Foundry Local (on-device)

Models run entirely on your device. The UI model picker shows every model in three states:

Status	Meaning
Loaded	Model is currently in memory — fastest start
Cached	Downloaded to disk, will load in seconds
Available	In the catalog, will download on first use

Cards show device type (GPU/CPU), size in MB, tool-calling support, publisher, and task type.

Microsoft Foundry (cloud)

Enter your Foundry endpoint URL and API key. Click List Models to browse your deployed models and click any entry to select it. The Deployment Name field is optional — leave blank to use the model name as the deployment identifier.

API Endpoints

Endpoint	Description
`GET /api/models/local`	Returns catalog models with live `loaded`/`cached`/`catalog` status
`GET /api/models/azure`	Lists deployed models from the configured Foundry endpoint
`GET /api/model-config`	Current provider, model, and endpoint settings
`POST /api/model-config`	Update provider/model settings (persists to `.env`)

Demo 1 — Maker-Checker PR Review

Pattern: Sequential — iterative review loop
Builder: SequentialBuilder
File: demos/maker_checker/agents.py

Agents

Agent	Role	Behaviour
Worker	Drafter	"You are a senior developer. Draft a concise PR review addressing code quality, bugs, and suggestions." Produces the initial review.
Reviewer	Quality gate	"You are a code review lead. Critique the draft against: correctness, clarity, actionability. Score 1-5." If score < 4, Worker revises.

Topology

Worker ──draft──► Reviewer ──score < 4?──► Worker (revision)
                     │
                  score ≥ 4
                     ▼
                    Done

When to use this pattern

Quality gates, approval steps, iterative document drafting, compliance checking, code review automation. Any task where the first attempt rarely meets the bar and a structured feedback loop improves quality.

Demo 2 — Hierarchical Research Brief

Pattern: Concurrent fan-out + Sequential synthesis
Builder: ConcurrentBuilder (specialists) + SequentialBuilder (synthesis)
File: demos/hierarchical_research/agents.py

Agents

Agent	Role	Behaviour
Manager	Decomposer	"Decompose the research topic into 2 specific sub-questions." Produces the task list.
Specialist_A	Technical researcher	"Research the technical aspects of the given topic." Runs in parallel with Specialist_B.
Specialist_B	Market researcher	"Research the market/business aspects of the given topic." Runs in parallel with Specialist_A.
Synthesizer	Report writer	"Combine the specialist reports into a cohesive 200-word brief." Runs after both specialists complete.

Topology

Manager ──► Specialist_A ──┐
        └──► Specialist_B ──┴──► Synthesizer ──► Done

When to use this pattern

Multi-source research, technical due diligence, market analysis, any task where independent sub-questions can be explored in parallel and merged into a single output.

Demo 3 — Hand-off Customer Support

Pattern: Explicit agent-to-agent control transfer
Builder: HandoffBuilder
File: demos/handoff_support/agents.py

Agents

Agent	Role	Behaviour
Triage	Classifier	"Classify customer issue as BILLING or TECH. Hand off to the right specialist." Emits a `handoff` event.
Billing	Billing specialist	"Handle billing inquiries: refunds, charges, payment methods." Receives control from Triage when issue is billing-related.
TechSupport	Tech specialist	"Handle technical issues: connectivity, errors, setup." Receives control from Triage when issue is technical.

Topology

Triage ──BILLING──► Billing ──► Done
       └──TECH────► TechSupport ──► Done

When to use this pattern

Helpdesk automation, intent-based routing, HR query routing, escalation pipelines. Use when control must transfer cleanly to a specialist without a shared conversation buffer.

Demo 4 — Network Brainstorm

Pattern: Peer group chat — all agents share one conversation thread
Builder: GroupChatBuilder
File: demos/network_brainstorm/agents.py

Agents

Agent	Role	Behaviour
Innovator	Creative thinker	"Propose bold, unconventional ideas. Build on others' suggestions."
Pragmatist	Feasibility evaluator	"Evaluate feasibility. Suggest practical implementations."
DevilsAdvocate	Critic	"Challenge assumptions. Identify risks and weaknesses."
Synthesizer	Integrator	"Find common ground. Synthesize the best elements into a plan."

Topology

┌─────────────────────────────────────────────┐
│  Innovator ◄──► Pragmatist                  │
│       ▲              ▲                       │
│       └───────────────┘                     │
│  DevilsAdvocate ◄──► Synthesizer            │
│  (all-to-all shared conversation)           │
└─────────────────────────────────────────────┘

When to use this pattern

Brainstorming, committee review, adversarial red-teaming, design critique. Best suited where emergent consensus is the goal and strict ordering would suppress contributions.

Demo 5 — Supervisor Router

Pattern: Classify then route
Builder: HandoffBuilder (autonomous mode)
File: demos/supervisor_router/agents.py

Agents

Agent	Role	Behaviour
Supervisor	Router	"Classify the task and select the best specialist." Emits a `handoff` to the matching expert.
CodeExpert	Code specialist	"Handle code generation, debugging, and review tasks."
DataExpert	Data specialist	"Handle data analysis, SQL queries, and visualization tasks."
DocExpert	Documentation specialist	"Handle documentation writing, editing, and structure tasks."

Topology

Supervisor ──Code──► CodeExpert ──► Done
           ──Data──► DataExpert ──► Done
           ──Docs──► DocExpert  ──► Done

When to use this pattern

Task dispatching, intent classification, tool selection, multi-skill agent systems. Separating the routing decision (Supervisor) from the execution (specialists) keeps each agent focused.

Demo 6 — Swarm + Auditor

Pattern: Concurrent generation + sequential audit + selection
Builder: ConcurrentBuilder (generators) + SequentialBuilder (audit/select)
File: demos/swarm_auditor/agents.py

Agents

Agent	Role	Behaviour
Generator_A	Creative proposer	"Propose bold, creative solutions." Runs concurrently with B and C.
Generator_B	Practical proposer	"Propose cost-effective, practical solutions." Runs concurrently with A and C.
Generator_C	Risk-averse proposer	"Propose safe, risk-averse solutions." Runs concurrently with A and B.
Auditor	Scorer	"Score each proposal on feasibility, impact, and cost." Runs sequentially after all generators.
Selector	Decision maker	"Pick the winning proposal based on auditor scores." Final sequential step.

Topology

Generator_A ──┐
Generator_B ──┴──► Auditor ──► Selector ──► Done
Generator_C ──┘
(concurrent)        (sequential evaluation phase)

When to use this pattern

A/B content generation, solution exploration, risk scoring, automated tenders, any scenario where generating diverse options in parallel and then evaluating them systematically yields better outcomes than picking a single path upfront.

Demo 7 — Magentic One Assessment

Pattern: Adaptive orchestration — manager decides flow dynamically
Builder: MagenticBuilder
File: demos/magentic_one/agents.py

Agents

Agent	Role	Behaviour
MagenticManager	Orchestrator	Reads the conversation state and decides which agent to invoke next, how many times, and when the task is complete. Not a participant in the output.
Researcher	Information gatherer	"Research the current state, trends, and real-world examples for the given topic."
Strategist	Strategy former	"Propose a recommended approach, adoption roadmap, and positioning for the given topic."
Critic	Risk assessor	"Identify the top risks, gaps, and mitigation recommendations."

Topology

         ┌──► Researcher ──┐
MagenticManager            ├──► MagenticManager (re-evaluates) ──► Done
         ├──► Strategist ──┤
         └──► Critic ──────┘
(dynamic routing — Manager decides order and repetition)

How Magentic One differs from other patterns

	Group Chat	Handoff	Magentic One
Turn order	Round-robin	Signal-based	LLM-decided
Re-invocation	Fixed rounds	No	Yes
Termination	Max rounds	Agent signals	Manager decides
Best for	Brainstorming	Clean routing	Open-ended tasks

When to use this pattern

Strategy assessment, feasibility analysis, research synthesis, open-ended tasks where the required steps, their order, and how many iterations are needed are not known in advance.

Summary Table

#	Demo	Pattern	Builder	Agents
1	Maker-Checker PR Review	Sequential loop	`SequentialBuilder`	Worker, Reviewer
2	Hierarchical Research Brief	Concurrent + Sequential	`ConcurrentBuilder` + `SequentialBuilder`	Manager, Specialist_A, Specialist_B, Synthesizer
3	Hand-off Customer Support	Handoff	`HandoffBuilder`	Triage, Billing, TechSupport
4	Network Brainstorm	Group Chat	`GroupChatBuilder`	Innovator, Pragmatist, DevilsAdvocate, Synthesizer
5	Supervisor Router	Classify + Route	`HandoffBuilder`	Supervisor, CodeExpert, DataExpert, DocExpert
6	Swarm + Auditor	Concurrent + Sequential	`ConcurrentBuilder` + `SequentialBuilder`	Generator_A, Generator_B, Generator_C, Auditor, Selector
7	Magentic One Assessment	Adaptive / Magentic	`MagenticBuilder`	MagenticManager, Researcher, Strategist, Critic

Pattern Decision Guide

Does the task have a fixed sequence of steps?
  └── YES → SequentialBuilder
  └── NO  → Can steps run in parallel?
              └── YES → ConcurrentBuilder (+ SequentialBuilder to merge)
              └── NO  → Does it need explicit routing to one specialist?
                          └── YES → HandoffBuilder
                          └── NO  → Are agents peers with equal contribution?
                                      └── YES → GroupChatBuilder
                                      └── NO  → Is the flow open-ended / unknown upfront?
                                                  └── YES → MagenticBuilder

FilesExpand file tree

agents.md

Latest commit

History

agents.md

File metadata and controls

Agent Patterns Demo Pack — Agent Reference

Model Configuration

Foundry Local (on-device)

Microsoft Foundry (cloud)

API Endpoints

Demo 1 — Maker-Checker PR Review

Agents

Topology

When to use this pattern

Demo 2 — Hierarchical Research Brief

Agents

Topology

When to use this pattern

Demo 3 — Hand-off Customer Support

Agents

Topology

When to use this pattern

Demo 4 — Network Brainstorm

Agents

Topology

When to use this pattern

Demo 5 — Supervisor Router

Agents

Topology

When to use this pattern

Demo 6 — Swarm + Auditor

Agents

Topology

When to use this pattern

Demo 7 — Magentic One Assessment

Agents

Topology

How Magentic One differs from other patterns

When to use this pattern

Summary Table

Pattern Decision Guide

Reference