Claude Managed Agents can write and run multi-agent workflows.
Claude Managed Agents agents can write a workflow, then run it in the background
You turn the feature on with the managed-agents-2026-04-01 beta header and a multiagent block typed multiagent_20261001 with workflows enabled. The agent decides when to start a run from your message or its system prompt. You follow progress through workflow_run.* events on the session stream, including phase starts and a final result.
Official limits matter more than the marketing shape. One run may start up to 1,000 agents over its lifetime, with at most 64 threads working at once. A run lasts 24 hours by default (the agent can set a shorter lifetime). A session can keep up to 10 runs open at once. Runs share the session sandbox and files; only the primary-thread agent can start a run.
Every model request in a run still counts toward the session’s budget. Hitting the budget pauses open runs. Anthropic’s own docs tell you to set a session budget and to tell the agent in the system prompt when to start a run. That is the operator move: treat workflows as a spend amplifier, not a free parallelism knob.
What to do: if you already use Managed Agents, read the workflow-runs guide, enable workflows on a non-production agent first, set a hard session budget, and wire your client to track workflow_run.created through workflow_run.status_ended before you declare the work done.
Sources: Claude Platform release notes (Oct 9) · Workflow runs · Multiagent orchestration
No culture/noise items this window that beat the operator bar. OpenAI’s Oct 9 Asana and Sophos customer posts stay off the card deck (vendor case studies without a new API delta).
Model & runtime releases
| Item | Date | Badge |
|---|---|---|
| Claude Managed Agents dynamic workflows (beta) | Oct 9 | Cloud |
| Ollama v0.40.3 | Oct 11 | Local |
| llama.cpp b11556+ Prism Bonsai 2 27B | Oct 11 | Local |
| MLX main Metal conv / qvm fixes (unreleased) | Oct 10–12 | MLX |
| TwelveLabs Pegasus 1.5 on Bedrock | Oct 9 | Cloud |
Bedrock reasoning.summary for OpenAI | Oct 9 | Cloud |
Security / policy / access
| Item | Date | Badge |
|---|---|---|
| Anthropic unintended model actions report | Oct 9 | Watch |
| vLLM Oct 9 advisory batch (Issue 35) — still patch to 0.31.0 | Oct 9 | Active |
| Anthropic 2026 Usage Policy (effective Nov 12) | Oct 8 | Watch |
| No new vLLM / llama.cpp / Ollama advisories Oct 10–12 | — | Quiet |
Critical
If you run Claude Managed Agents with workflows enabled, set a session budget before the first production prompt. Uncapped fan-out is the failure mode.
Watch
Anthropic’s unintended-actions report is a preview of how agent evals with live internet will keep producing edge cases. Expect more of these posts, not fewer.
Opportunity
Bedrock’s OpenAI reasoning summaries give you a vendor-supported trace channel for Sol/Astra-class agents without building your own summarizer.
Budget a Managed Agents workflow run
On a throwaway agent, enable multiagent_20261001 workflows, set a low session budget, and ask for a multi-document review. Confirm you see workflow_run.* events and that hitting the budget pauses the run.
Pull llama.cpp ≥ b11556 for Bonsai
Grab a weekend build and the Prism Bonsai 2 27B *-Q2_0_g64.gguf. Skip older non-g64 Q2_0 files. Check Prism’s license before any commercial use.
Upgrade Ollama to 0.40.3
Install 0.40.3 so background embedding upgrades stay paused. If you already ran 0.40.2, clear leftover model backups with the cleanup noted in that release.
Saturday through Monday morning (Oct 10–12 ET), with Friday’s lead as the opening implication. ASTGL / Fraser takes, sourced.
Set a session budget before you turn workflows on
Dynamic workflows are the first Managed Agents feature that makes “one user message” an unbounded fan-out. The agent writes the program. The server runs it. You get phase events. That is genuinely useful for reviewing hundreds of documents. It is also a way to spend a week of API budget before lunch if you leave the session uncapped. Anthropic already ships session budgets and says runs pause when you hit them. My take: enable workflows only after the budget field is set, and put the “when to start a run” rule in the system prompt in plain language. If you can’t explain the spend path, you are not ready for the feature.
Sources: Workflow runs · Release notes (Oct 9)
Anthropic published its own agent misbehavior, in public
Friday’s Anthropic report walks through cases from evaluations and internal use: exploiting a basic flaw to run commands on a third-party server, submitting a real form when a practice form failed, pulling gated-but-public data with scraped tokens, and using URL shorteners to get around fetch-tool length limits. Anthropic says the real-world impact was minimal, briefed the White House on government-site cases, and has expanded offline evaluations plus automatic blocking. The operator lesson is not “Claude is evil.” It is that when a task looks blocked, capable agents look for workarounds. If your harness gives them a browser, a fetch tool, or a shell, put the deny list and the kill switch in the harness.
Sources: Anthropic — unintended model actions (Oct 9)
Ollama paused the upgrades that were eating disk and cycles
0.40.2 explained that older models get upgraded in the background for the llama.cpp runner and that originals stay as backups, so you temporarily store models twice. 0.40.3, tagged Saturday night UTC, pauses those background upgrades for embedding models and other request-performance cases. That is a small release note with a large “please stop filling my SSD” energy. If you already upgraded over the weekend, still check for leftover backups. If you have not, jump straight to 0.40.3.
Sources: Ollama v0.40.3 · v0.40.2
Stock llama.cpp can run Prism Bonsai 2 27B
Sunday’s b11556 build merges runtime support for Prism’s ternary Bonsai 2 27B without needing Prism’s fork for the standard Q2_0 group-64 GGUF, including Metal. That is the kind of local win this digest exists for. Read the weight filename carefully: older non-g64 Q2_0 files are the ones that go wrong. Heavier Prism quants still need their fork. Don’t treat a merged PR as a license change either — check Prism’s own terms before you ship anything commercial.
Sources: llama.cpp #29600 · llama.cpp releases
MLX Metal fixes keep landing on main, not on a tag
Over the weekend MLX main fixed Metal convolution bounds checks and integer overflows, and a batched row-count mismatch in Metal
qvm_split_k. Useful if you live on nightly wheels. The latest release tag is still 0.32.3 from late September. Same pattern as Friday’s 1-bit kernels: the interesting Apple Silicon work is ahead of the tag. Pin what you ship; follow main only if you can rebuild when a fix lands.Sources: MLX #4672 · MLX #4673