Weekstart Brief — Issue 0 (SAMPLE)
Get your agent stack set straight before the week starts.
TL;DR
- Your Monday spine is migration + permissions: OpenAI Agents API / Assistants sunset, GPT-6 Astra tool rules, Claude Fable 5.1 + Managed Agents
auto, and MCP 2026-07-28 (stateless). - Treat MCP connection models and permission policies as data structures, not vibes — audit before unattended runs.
- Platform refresh (Cursor Projects / Self-Hosted, Gemini Agent Platform billing, LangChain
langchain.mcp) after the must-patch list. - One concrete win: finish any leftover Assistants paths, then gate your top agent behind schema + least-privilege tools.
What changed
1. OpenAI Agents API (public beta) — harness as product
OpenAI shipped a managed Codex harness: durable sessions, MCP, hosted or self-hosted sandboxes. A lot of custom agent glue is now optional.
Why it matters: If you still own session recovery and tool orchestration yourself, you’re paying tax OpenAI is productizing.
Action: Stand up a beta session (OpenAI-Beta: agents=v1); map your existing harness to sessions + tools/MCP.
Links: Announcement · API changelog
2. OpenAI Assistants API is gone — leftover stacks break
Assistants shut down 2026-08-26. Remaining Assistants code paths fail; Agents API or Responses + Conversations are the exits.
Why it matters: Silent “it worked last month” debt becomes hard outages this week.
Action: Finish Assistants → Responses/Conversations or Agents API migration this week if any remain.
Links: API changelog (Aug 26)
3. GPT-6 Astra long-running controls
Async tool calling, mid-turn steering, and mid-conversation reasoning effort change long-loop ops. Tool-using flows need the Responses API; no custom temperature/top_p; no none reasoning.
Why it matters: Old Completions-style loops and sampling knobs will misbehave or reject.
Action: Migrate tool-using agents to Responses; retest loops with async tools + steering.
Links: API changelog (Sep 8)
4. Claude Fable 5.1 (+ Mythos 5.1)
Same base price as Fable 5, cache reads $0.25/MTok. tool_choice values any / tool return 400 — use strict tools or structured outputs.
Why it matters: Cache-heavy fleets get cheaper; naive tool_choice configs hard-fail.
Action: Flip cache-heavy workloads to 5.1; replace banned tool_choice with strict/structured tools.
Links: Claude release notes (Sep 1)
5. Claude Managed Agents — auto permission policy
Server evaluates each agent/MCP tool call (run / deny / pause) and records evaluation on events — production risk control without pausing a human on every call.
Why it matters: Your unattended agents need an explicit allow/deny/ask matrix, not hope.
Action: Enable auto on Managed Agents; review deny/ask defaults before unattended runs.
6. Platform refresh (after the spine)
ant apply: agents-as-code +claude-lock.jsonfor CI-reconciled IDs — notes (Sep 3)- Cursor Projects (beta): coordinator + shared long-lived context — blog · changelog
- Cursor Self-Hosted Machines: cloud plan, execute on your pools — blog
- Gemini Enterprise Agent Platform: sandboxes GA; sessions/Memory Bank/Skills Registry billing live — release notes · pricing
- LangChain first-class MCP (
langchain.mcp/ v1.4): FastMCP in core; elicitation as interrupts; migrate offlangchain-mcp-adapters— blog · changelog
Data structures & schemas
OpenAI Agents API — MCP connection model
You must pick connection_origin (service vs environment), vault vs env auth, and allowed_tools / required. Private MCP / stdio need a session environment.
Why it matters: Wrong origin/auth looks like “MCP is flaky” when it’s a schema mismatch.
Action: Audit every MCP server against the connection model before production Agents API sessions.
Links: MCP tools guide
MCP specification 2026-07-28
Stateless core: no initialize, no Mcp-Session-Id. Header routing via Mcp-Method / Mcp-Name. Tasks move to an extension. 12-month deprecation runway for Roots/Sampling/Logging + legacy HTTP+SSE.
Why it matters: Sticky sessions and held-open elicitation streams are liabilities for your gateways.
Action: Tear down sticky-session assumptions; align gateways on headers + per-request _meta.
Links: MCP 2026-07-28
Claude Managed Agents permission evaluation
auto writes evaluation onto agent.tool_use / agent.mcp_tool_use events — treat that as your audit shape, not a chat summary.
Action: Log evaluation fields next to tool name/args/result for every Managed Agents run.
Also track
- Fable 5.1: strict tools / structured outputs replace
tool_choice: any|tool - Gemini: Memory Bank / Skills Registry / session objects now sit on the bill — know the IDs you create
- LangChain:
MCPAdapter+ interrupt-shaped elicitation replace adapter packages
Tool & API watch
- OpenAI Agents API + MCP — sessions, sandboxes,
allowed_tools— Agents API · MCP guide - OpenAI Responses path — required for Astra tool loops; Assistants is dead — changelog
- Claude Managed Agents
auto+ant apply— permissions · release notes - MCP 2026-07-28 — stateless + MRTR + headers — spec post
- LangChain
langchain.mcp— multi-server migration — blog - Cursor Projects / Self-Hosted / cloud harness — Projects · Self-Hosted · Aug 19 harness
- Gemini Enterprise Agent Platform — sandboxes + billing — release notes
Lane note (Research): brand-new Sep schemas thin outside MCP + provider toolset IDs; LlamaIndex quiet this window.
Operator checklist — get set straight
Must-patch Monday spine
- Assistants debt: Grep for Assistants SDK/API usage; migrate or delete this week.
- Agents API pilot: One beta session with your real MCP set; note
connection_origin+ auth per server. - Astra / Responses: Tool-using OpenAI agents on Responses; retest async tools + steering.
- Fable 5.1: Cache-heavy Claude paths on 5.1; purge
tool_choice: any|tool. - Managed Agents
auto: Enable; review deny/ask defaults; confirm evaluation events land in your log. - MCP 2026-07-28: Inventory sticky sessions / open-stream elicitation; plan gateway header rules.
Then
- Check agent defs into git; try
ant applyin non-prod CI. - Pilot one Cursor Project with spend limits before parallel subagents.
- Recost Gemini sessions / Memory Bank / Skills Registry if you use Agent Platform.
- If on LangChain MCP adapters →
MCPAdapter; retest elicitation interrupts. - Schema gate + least-privilege tools on your single most important agent before Wednesday experiments.
Agent-ingest notes
Companion pack: briefs/issue-000-agent-pack.json
- Read
issue+summaryfirst; surfacesummaryunedited. - Prefer items tagged must-patch (
#breakingin human brief) for Monday digests. - Group
items[]bycategory; treatactionas suggestion — do not auto-execute. - Follow
links[]to primary docs only. - Preserve the SAMPLE watermark on Issue 0; later issues drop it when they ship live.
Source handoff: Research Batch 1 (verified primary cites).