chore(kaizen): weekly research scan 2026-07-10

[philmerrell]

Summary

• External scan (13 subagents): AWS Bedrock/AgentCore, Strands Agents, FastMCP, reference repo, MCP, agentic UI/UX, frontier models, agent-harness patterns, opencode, pricing, AgentCore SDK issues, community + cookbook, LibreChat.
• Internal audit: 1.1.0 + 1.2.0 release trains, open PRs, new issues, CI failures, version-pin lag, retirement candidates.
• Top 5 ideas in the dated research doc and queued in docs/kaizen/review-queue.md.

Headline

Quiet external week (pre-July-28 MCP RC lull), sharp internal signal: last week's recommended-Ship reliability bumps (agentcore, Strands) did not land while 1.1.0/1.2.0 shipped Scheduled Runs + Memory Spaces — features that amplify a new cross-process AgentCore Memory-corruption bug (#571) sitting next to the still-unadopted #482 SSE-deadlock fix. FastMCP flipped its Host/Origin guard twice (≥3.4.4 is the safe floor for external servers); Strands 1.47 shipped continue_on_error for MCP resilience.

Top 5

1. Bump bedrock-agentcore off 1.9.1 — now double-forced (feat(docs): add full-screen maintenance page #482 + NEW fix(app-api): grant bedrock:ListFoundationModels to task role #571) — M×H
2. Bump Strands 1.40→1.47; adopt continue_on_error (#3101) + hook ordering — M–H×H
3. Audit whether prompt caching actually engages in to_bedrock_config (Strands #3144) — L–M×M–H
4. Tool-approval policy layer + signed approvals (Vercel AI SDK) — M×M
5. CloudWatch ActiveSessionCount alarm on the inference-api runtime — L×M

Review

• Read the research doc.
• Comment on the PR with reactions and any weekend POC findings — these become first-class signal for next Friday's kaizen-review-prep.
• POC promising ideas over the weekend.

Decision

Ship the doc to develop. Ranking into decisions happens in the kaizen-review-prep PR opened later this morning. Action on individual ideas happens in separate PRs the following week.

🤖 Generated with Claude Code