docs: add Kiro Crew to landscape analysis

[ralphbean]

Summary

• Adds a Kiro Crew entry to the landscape analysis (AWS Kiro's open-sourced multi-session orchestration layer, announced 2026-08-04): its ACP-based observable sub-agent delegation, Memory/Lessons/Skills stack, "Apps," and sandbox security model, each compared to fullsend's positions.
• Cross-references Kiro Crew from cross-run-memory.md — its Memory/Lessons/Skills split is a real-world precedent for the trust-tier distinction that doc already draws, and its automatic (vs. review-gated) promotion model sharpens the open question there.
• Cross-references Kiro Crew from agent-architecture.md — its live observability of planning/delegation/approvals is worth borrowing, but its coordination model is the same single-coordinator topology as MetaGPT/CrewAI, and it's session-scoped rather than repo-scoped.

Test plan

• make lint passes on staged changes (markdown links, docs link scope)

[fullsend-ai-review]

🤖 Finished Review · ✅ Success · Started 4:08 PM UTC · Completed 4:23 PM UTC
Commit: e6414bd · View workflow run →

[qodo-code-review]

PR Summary by Qodo

Docs: add Kiro Crew to landscape analysis and cross-reference key concepts

📝 Documentation 🕐 10-20 Minutes

AI Description

• Add a new Kiro Crew entry to the agent-orchestration landscape, highlighting architecture, memory
 tiers, and security model.
• Cross-reference Kiro Crew from agent-architecture and cross-run-memory as concrete precedents for
 observability and trust-tiered memory.

Diagram

graph TD
  M["docs/problems/cross-run-memory.md"] --> L["docs/landscape.md (Kiro Crew)"] --> K{{"Kiro Crew"}}
  A["docs/problems/agent-architecture.md"] --> L

Loading

High-Level Assessment

The approach (add a landscape entry, then cross-link from relevant problem docs) is the right fit for the existing doc structure. A separate standalone “case study” page was a possible alternative, but would add navigation/maintenance overhead without clear benefit for a single new ecosystem datapoint.

Files changed (3) +20 / -0

Documentation (3) +20 / -0

landscape.mdAdd Kiro Crew landscape entry (architecture, memory tiers, security) +14/-0

Add Kiro Crew landscape entry (architecture, memory tiers, security)

• Introduces a new “Kiro Crew” section covering its ACP-based observable delegation model, Memory/Lessons/Skills stack, scheduled “Apps,” and sandbox/security posture. Explicitly compares these design choices to fullsend’s trust boundaries and review-gating questions, and links to related internal docs.

docs/landscape.md

agent-architecture.mdReference Kiro Crew as an observable delegation example +2/-0

Reference Kiro Crew as an observable delegation example

• Adds a new bullet describing Kiro Crew’s live observability of planning/delegation/tool use via ACP. Positions it as a star-topology, single-coordinator pattern (like MetaGPT/CrewAI) and contrasts it with fullsend’s repo-scoped authorization and zero-trust composition goals.

docs/problems/agent-architecture.md

cross-run-memory.mdAdd industry precedent section citing Kiro Crew’s memory tiers +4/-0

Add industry precedent section citing Kiro Crew’s memory tiers

• Adds an “Industry precedent” section pointing to Kiro Crew’s Memory/Lessons/Skills split as corroboration for the document’s trust-tier separation. Highlights the key difference: Kiro’s automatic promotion with post-hoc inspection vs. fullsend’s open question of pre-promotion review gating across roles/agents.

docs/problems/cross-run-memory.md

[qodo-code-review]

Code Review by Qodo

🐞 Bugs (0) 📘 Rule violations (1) 📜 Skill insights (0)

1. cross-run-memory.md lacks options 📘 Rule violation ⌂ Architecture

Description

The updated problem document does not document at least two distinct solution options with clearly
labeled trade-offs, which is required for core problem docs. This makes it harder to evaluate
alternatives and can read like an implicit single-track direction.

Code

docs/problems/cross-run-memory.md[R65-68]

+## Industry precedent
+
+[Kiro Crew](../landscape.md#kiro-crew) ships a three-tier stack — Memory (preferences/project context), Lessons (corrections promoted into "durable lessons"), Skills (repeated patterns promoted into inspectable, editable artifacts) — that lands close to the stable/tactical/self-assessment split above, independently arrived at. It is a useful existence proof that the split is practically necessary rather than a theoretical nicety. But its promotion model differs from what this document is asking: Kiro Crew promotes automatically and makes the result inspectable after the fact ("you decide what your crew carries forward" is a review step that happens post hoc, by the same user who is already trusting the tool). The open question here is whether promotion needs to be gated *before* a lesson can reach a different agent — particularly across roles, per [What must be separated](#what-must-be-separated) above — rather than left to after-the-fact inspection by whoever is already running the crew.
+

Relevance

● Weak

Closest precedent: same file had “add options + trade-offs” request rejected by team.

PR-#3065
PR-#3063

ⓘ Recommendations generated based on similar findings in past PRs

Evidence

PR Compliance ID 1062035 requires that any new or modified problem doc under docs/problems/
explicitly presents at least two solution options with trade-offs. The updated cross-run-memory.md
adds context (## Industry precedent) but the document content around this change remains
contextual framing rather than a structured options-and-tradeoffs comparison.

Rule 1062035: Problem docs must present multiple options with trade-offs, not a single prescribed solution
docs/problems/cross-run-memory.md[65-68]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
`docs/problems/cross-run-memory.md` is a core problem document and must include at least two distinct solution options/approaches, each with clearly labeled trade-offs (pros/cons/risks), rather than only background/constraints/open questions.

## Issue Context
This PR adds new material to the problem doc (`## Industry precedent`) but does not add an options + trade-offs comparison section.

## Fix Focus Areas
- docs/problems/cross-run-memory.md[65-68]

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---

Context used

✅ Compliance rules (platform): 54 rules

✅ Skills: writing-user-docs, writing-adrs

_{To customize comments, go to the Qodo configuration screen, or learn more in the docs.}

[github-actions]

Site preview

Preview: https://bc96a5c8-site.fullsend-ai.workers.dev

Commit: fa0ef8d6d311441ccc7116d9d068daa30976a885

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[fullsend-ai-review]

Looks good to me

Previous run

Review

Findings

Medium

• [internal-consistency] docs/problems/agent-architecture.md:210 — The new paragraph states Kiro Crew is “one developer’s crew, running on their machine or Kiro’s hosted execution.” However, the Kiro Crew entry added in the same PR to docs/landscape.md explicitly states: “There is no AWS/Kiro-hosted execution tier” and describes the deployment model as “Local-first and self-hosted, not a managed service.” This is a direct internal contradiction within the same PR.
  Remediation: Change “running on their machine or Kiro’s hosted execution” to something consistent with the landscape entry, e.g., “running on their own machine or a self-hosted server” or “running on their own machine or a remote host they control.”

Previous run (2)

Review

Findings

Low

• [internal-consistency] docs/problems/cross-run-memory.md — The Skills tier description says "inspectable, editable artifacts" but docs/landscape.md describes the same concept as "inspectable, editable, removable artifacts." The word "removable" is missing from the cross-run-memory.md version.
  Remediation: Add "removable" to match: Skills (repeated patterns promoted into inspectable, editable, removable artifacts).

• [internal-consistency] docs/problems/cross-run-memory.md — The mapping between Kiro Crew's three-tier stack and the document's "stable/tactical/self-assessment" split is imprecise — Lessons (corrections from past interactions) maps more closely to "agent self-assessment" than to "recent operational state." The text hedges with "close" rather than claiming exact correspondence, so this may be intentional.

---

Labels: PR adds a new entry to the landscape analysis document and cross-references it from two problem docs

[fullsend-ai-review]

🤖 Finished Review · ✅ Success · Started 4:28 PM UTC · Completed 4:43 PM UTC
Commit: 1b997b3 · View workflow run →

[waynesun09]

[MEDIUM] New Kiro Crew entry postdates the doc's "Conducted" staleness marker without updating it

docs/landscape.md line 5 opens with > **Conducted: 2026-03-06.** ... Treat the specific tool capabilities and pricing as potentially stale. This PR adds a new Kiro Crew section describing an announcement dated 2026-08-04 (five months after the stated survey date) without touching that header (confirmed: gh pr diff shows no changes to the "Conducted" line). The document's own freshness contract is now internally inconsistent — a reader has no way to tell from the header that at least one entry is current as of August, not March.

Suggestion: Either bump the "Conducted:" date if it represents "last updated," or add a short note next to it acknowledging incremental additions post-dating the original survey (e.g., "entries may be added incrementally after this date; see git blame for per-entry recency").

(Note: line 5 falls outside this PR's diff hunks, so this couldn't be attached as an inline comment.)

[fullsend-ai-review]

🤖 Review · ⚠️ Cancelled · Started 6:39 PM UTC · Ended 6:41 PM UTC
Commit: 5507333 · View workflow run →

[fullsend-ai-review]

🤖 Review · ⚠️ Cancelled · Started 6:42 PM UTC · Ended 6:54 PM UTC
Commit: 1b997b3 · View workflow run →

[waynesun09]

Approving — no blocking issues. The stale 'Conducted' marker I flagged is resolved. One open MEDIUM fast-follow worth a look: agent-architecture.md's new paragraph says Kiro Crew runs on "their machine or Kiro's hosted execution," which contradicts this same PR's landscape.md entry stating there is no AWS/Kiro-hosted execution tier — a direct internal inconsistency within the PR's own diff, non-blocking but worth reconciling. CI green.

[fullsend-ai-review]

🤖 Finished Review · ✅ Success · Started 6:55 PM UTC · Completed 7:09 PM UTC
Commit: fa0ef8d · View workflow run →

[waynesun09]

[MEDIUM] "There is no AWS/Kiro-hosted execution tier" overstates certainty and elides the credit-metered kiro-cli dependency.

The Deployment model paragraph states as settled fact: "There is no AWS/Kiro-hosted execution tier." Two issues compound here: (1) the only cited source text (KiroCrew docs) says the Gateway/orchestration "does not require a Kiro Crew-hosted control plane" — i.e. documents that a hosted control plane isn't required, not that AWS has no hosted tier at all or won't ship one (Kiro IDE itself has Enterprise/managed-update features per kiro.dev/docs, so this isn't implausible). (2) More concretely, the actual inference backend is not local-only: Agent Sessions run kiro-cli, which per independent coverage (InfoWorld, SiliconANGLE) requires a Kiro account/sign-in and is metered by credits (subscription tiers, credit-card-gated overages) — a real, load-bearing dependency on Kiro-hosted billing/inference. The blanket "no hosted execution tier" framing conflates 'the Gateway/orchestrator is self-hosted' (true) with 'there is no vendor-hosted dependency anywhere in the stack' (not established, and likely false for the inference layer) — and this framing is what the downstream single-tenant argument and the closing "no described mechanism for one Gateway to safely serve multiple teams" claim are built on.

Suggestion: Qualify the claim to separate orchestration from inference, e.g.: "The Gateway/orchestration layer is local-first and self-hosted — no AWS/Kiro-hosted orchestrator ('does not require a Kiro Crew-hosted control plane'). The inference backend is different: Agent Sessions drive the proprietary, credit-metered kiro-cli, which requires a Kiro account/sign-in, so the deployment is self-hosted for coordination but vendor-dependent for execution." This also sharpens the "Relevance to fullsend" argument about dependency coupling rather than weakening it.

[waynesun09]

[MEDIUM] Unsourced internal-codename claim ("MeshClaw") stated as settled fact with no citation.

'Kiro Crew, open-sourced 2026-08-04 (started internally at Amazon as "MeshClaw")' asserts a specific internal pre-launch codename as fact. Unlike the file's other granular claims (e.g. other entries' precise star/commit counts), which are directly checkable against a public GitHub repo, an internal-project codename from before open-sourcing is not verifiable from the linked announcement/repo/docs URLs at the top of the section, and no inline citation is given for it specifically. If it came from the blog post, a direct quote/footnote would remove the ambiguity; if from a secondary source (tweet, former-employee comment), that source should be named. As written it reads as an unverified detail asserted with the same confidence as the sourced material around it.

Suggestion: Either add a direct citation/quote for the "MeshClaw" claim (a link to where AWS or a credible reporter states this), or soften it to "reportedly started internally at Amazon" with an inline source, or drop the codename if it can't be pinned to a citable source.

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 8:03 PM UTC · Completed 8:11 PM UTC
Commit: fa0ef8d · View workflow run →

[fullsend-ai-retro]

Retro: PR #5941 — docs: add Kiro Crew to landscape analysis

Workflow shape: Human-authored docs PR (3 files, +20 lines) → 3 successful review agent runs + 2 cancelled → human review → merge. No triage, code, or fix agent runs.

Timeline

1. 16:07 UTC — ralphbean opens PR with initial commit e6414bd.
2. 16:08–16:23 — Review agent run 1 finds 2 Low-severity internal consistency issues and approves: (a) missing word "removable" in cross-run-memory.md vs landscape.md, (b) imprecise tier mapping (intentional hedge by author).
3. ~16:27 — Author self-corrects landscape.md deployment model description (commit 1b997b3), inadvertently creating a new contradiction with agent-architecture.md.
4. 16:28–16:43 — Review agent run 2 correctly catches the newly introduced Medium-severity contradiction: agent-architecture.md says "Kiro's hosted execution" while the corrected landscape.md says "no AWS/Kiro-hosted execution tier."
5. 18:33 — Human reviewer waynesun09 flags that the "Conducted: 2026-03-06" staleness marker in landscape.md was not updated despite adding content dated 2026-08-04 — an out-of-diff metadata inconsistency.
6. 18:47 — waynesun09 approves, noting the agent-architecture contradiction as a non-blocking fast-follow.
7. ~18:53 — Author pushes fix commit fa0ef8d addressing the "removable" word and the hosted-execution contradiction.
8. 18:55–19:09 — Review agent run 3 gives clean approval.
9. 19:55 — waynesun09 posts 2 additional substantive inline comments (unsourced "MeshClaw" codename and orchestration-vs-inference conflation). These remain unresolved.
10. 20:01 — PR merged.

What went well

Review agent excelled at internal consistency. The agent caught a real contradiction between two files in the same PR (the hosted-execution discrepancy) and a missing-word inconsistency — both were fixed by the author. Notably, when the author's self-correction on commit 2 introduced a new cross-file inconsistency, the re-review correctly detected it. This is exactly the kind of mechanical consistency check where agents add reliable value.

Human and agent reviews were complementary. The agent focused on within-diff consistency; the human reviewer focused on out-of-diff metadata (stale "Conducted" date), factual accuracy (unsourced claims), and epistemic rigor (conflating orchestration with inference). Neither fully covered the other's findings.

Cancelled runs handled cleanly. Two review runs (31035674765, 31035910561) were cancelled by concurrency groups — the debounce mechanism working as designed.

Existing issues receiving supporting evidence

• Classify file-internal consistency errors (numbering, cross-references) as Medium, not Low #986 (classify file-internal consistency errors as Medium, not Low): The "removable" word omission was classified as Low but was worth fixing — supporting evidence that consistency errors between files in the same PR may warrant Medium.
• Review agent should flag out-of-diff documentation consistency findings as follow-up candidates #3695 (flag out-of-diff documentation consistency findings): The stale "Conducted" date is exactly the class of out-of-diff finding this issue targets.
• Track autonomy readiness: review agent outperformed human review on mechanical consistency in documentation repo #3377 / Track autonomy readiness: review agent outperforms human review on mechanical consistency checks in konflux-ci/architecture PR #368 #4995 (autonomy readiness for mechanical consistency in docs): This PR adds another data point — the review agent's consistency findings were actionable and correct, matching or exceeding the human reviewer's initial focus on that dimension.

Proposal filed

1. docs-review skill should detect stale document metadata date markers — see proposal below.

Proposals filed

• docs-review skill should detect stale document metadata date markers when content is added (in fullsend-ai/agents)