The old bottleneck was humans reviewing agent output; the new one is agents coordinating with each other. Claude Managed Agents flows hand that coordination to Claude itself, which plans phases, fans out subagents, and reports every step through a durable event stream you can interrupt.
What Claude Managed Agents Flows Are
Claude Managed Agents flows are a feature of Claude Managed Agents, Anthropic API for letting the vendor run the full agent life cycle, in which Claude designs and executes its own multi-agent topology instead of you wiring one by hand. Anthropic CJ Avilla and Harrison Stall presented them at AI Engineer NYC in October 2026, and Anthropic shipped the underlying multi-agent configuration type (multi-agent 2026-10-01) that same week. You hand the lead agent a goal and a budget; it plans phases, fans out subagents, and reports every step through a durable event stream.
The product itself launched in April 2026 as an API surface where you define an agent (a bundle of skills, tools, and MCP servers), instantiate it in a session, and let it run on a managed environment. Flows add self-orchestration on top: the lead agent decides how many subagents to spawn, spins them up and down per phase, and can use preconfigured specialist subagents you supply in the agent config.
Why a Single Agent Breaks Down on Big Tasks
A single agent fails on large tasks for three reasons the speakers named from their own experience: agentic laziness, self-preferential bias, and goal drift.
- Agentic laziness. Asked to check 50 services for an outage root cause, one agent that has seen 35 healthy services may conclude the rest are fine and stop early to save tokens.
- Self-preferential bias. An agent asked to review code or a poem it just wrote tends to approve its own work. Splitting generator and reviewer into isolated contexts produces better results.
- Goal drift. Constraints like "don't touch production" can get compacted away as the context window fills over a long run.
The speakers' framing echoes old organizational lessons. Ken Thompson built Unix alone in 1969 with full conceptual integrity; IBM's OS/360 needed roughly 5,000 person-years, and Fred Brooks recorded the coordination lessons in The Mythical Man-Month in 1975. When Linus Torvalds hit his limit reviewing patches in 1998, the Linux project eventually moved to layered trusted maintainers. Flows apply the same idea to agents: isolate context, distribute scope, escalate only exceptions.
How a Flow Runs, Step by Step
Running a flow through the API takes a handful of moves, and most of the orchestration work happens server-side:
- Create an agent resource, passing the multi-agent configuration type and a system prompt, plus optional preconfigured subagents, skills, MCP servers, and custom tools.
- Create or reuse an environment, the computer Claude operates on; you can bring your own self-hosted sandbox.
- Start a session from that agent and environment, and send a first event such as "run a workflow."
- Set a budget before kickoff so the session cannot burn unlimited tokens.
- Watch the single real-time event stream, and interrupt with a user.interrupt event if the run heads the wrong direction.
Two structural limits matter for planning. The current nesting depth is one layer: a lead agent with its own subagents, and Anthropic stated deeper mega-fan-outs are not supported yet while safety is hardened. The stated ceiling is up to 1,000 agents per flow, and a lead agent can chain multiple flows. Stall reported running a 1,000-agent flow for over 85 hours, producing roughly 1,000 commits toward a Unix-compatible operating system with a Python interpreter; that figure is a first-hand demonstration, not an independent benchmark. Subagents do not message each other directly; they share a file system, and every hand-off appears as thread message events you can inspect.
What Flows Cost and Where to Spend Compute
Flows cost more than a single-agent run, and the trade is coverage: in the speakers' own watchlist-scanner comparison, the flow found substantially more detail than one agent scanning every ticker alone. That comparison is vendor-reported from their demo, so treat the magnitude as indicative rather than measured.
You can allocate models per phase. In the stock watchlist scanner demo, cheap, fast Haiku agents scan the web two per ticker, while pricier Fable agents do the in-depth review at two or three per ticker. Prompts can expand or contract investment per phase, and Anthropic recommends validating important flows against an eval set. A practical optimization loop also ships in Claude Code, Anthropic terminal-based coding agent: the hill climb skill iteratively improves your prompt, and full flow transcripts can be fed back to Claude to find over- or under-worked phases. A related product, Dreaming, crystallizes memory stores across repeated sessions.
Bug Hunter and Explainer Studio Templates
Two shipped templates show the closed-loop patterns flows are good at. Both are available as console templates and installable in Claude Code via the Claude skill.
Bug Hunter for PR Review
Bug Hunter reviews a pull request by splitting the code into slices, giving each slice an isolated reviewer agent, then running a second-phase reviewer that agrees or disagrees with each flagged problem. A prover phase can clone the repo and attempt to reproduce the bug before findings are posted as prioritized inline PR comments. The value is coverage no single agent matches, with humans still making the final call.
Explainer Studio for Video
Anthropic uses a flow to generate many of its own animated product explainers. Research, fact-checking, and a design brief feed a production phase that builds and animates scenes, followed by voiceover and sound phases. Stall estimates a video that once took around 20 hours of animation work now starts from one small prompt, a first-hand estimate, not a measured benchmark.
Running 1,000 Agents Safely
Anthropic safety model for flows rests on three developer jobs: declare, contain, detect. Declare the agent's persona, tools, and permission policy. Contain it with budgets, tool restrictions, and network lockdowns in the environment configuration, such as allowing only api.twilio.com. Detect problems through the durable event stream, which logs every tool call, and by pointing a reviewer agent at past session transcripts.
Credentials Never Touch the Sandbox
The brain (inference loop and session state) is separated from the hands (the environment where actions execute). Containers receive fake placeholder secrets that fail outside the platform; the real credential is substituted only at the egress proxy, matched by destination host and injection location such as a bearer token. This makes prompting a key out of the sandbox structurally impossible, as presented by Anthropic.
Auto Mode and MCP Controls
Anthropic cites research that humans approve roughly 97% of permission prompts, so approval fatigue is the weak link; switching the permission policy to auto mode lets the model evaluate each tool call instead. MCP toolsets can be enabled tool by tool, guarding against a compromised server adding an exfiltration tool. For private internal MCP servers, MCP tunnels over Cloudflare give access without exposing anything to the open internet. These are platform controls; they do not by themselves constitute a compliance certification for regulated deployments.
Flows or Claude Code Dynamic Workflows?
Choose Claude Code dynamic workflows for interactive, one-off work at your terminal, and Claude Managed Agents flows when you need a repeatable, schedulable, API-embedded workflow. Anthropic generalized the ultracode methodology from Claude Code into the flows API, so the capability is similar but the integration surface differs.
| Dimension | Claude Code dynamic workflows | Managed Agents flows |
|---|---|---|
| Interface | Terminal, interactive | API, embeddable in your product |
| Best fit | One-off personal tasks | Scheduled, webhook-driven, CI runs |
| Orchestration | Claude self-orchestrates | Claude self-orchestrates via API config |
| Safety controls | Local permission modes | Budgets, vaults, egress lockdown, event stream |
| Environment | Your machine or cloud | Hosted or self-hosted sandbox |
A Kanban card dragged to a new column, a cron job before market open, or a webhook from a PR are the canonical flow triggers. The console dashboard itself is built on the public API, so a custom monitoring UI is a build-it-yourself option.
FAQ
- How many agents can one flow run? Anthropic states a flow can use up to 1,000 agents, with one layer of nesting under the lead agent today. A lead agent can also chain flows sequentially for larger projects.
- Do subagents talk to each other directly? There is no API to force subagent-to-subagent messages. They share a file system and can coordinate through it, and every message hand-off is visible as a thread event.
- What happens when a flow hits its budget? The session emits a budget-reached event and pauses. You can raise the budget mid-run and the flow resumes from where it stopped, keeping completed work.
- Can a human step in mid-flow? Yes. You can interrupt a session via the user.interrupt event and reprompt, or define a custom tool the agent calls whenever it needs human feedback.
- Are flows compatible with self-hosted sandboxes? The speakers said flows work out of the gate with self-hosted sandboxes, though a few platform features are not fully supported there. The environment type you pass at session creation changes accordingly.
From Conference Talk to Written Reference
The core insight of this session is that orchestration, not model capability, is the bottleneck: the same coordination lessons Brooks wrote down in 1975 now apply to a lead agent managing a thousand subagents. That kind of layered argument lives in a 102-minute talk, and most people will never watch all of it. If you have workshops, interviews, or technical sessions like this one sitting on YouTube, you can turn a video into a structured, source-linked article with Skala Blog: paste the URL, get the transcription, and edit a draft that preserves the sequence and the caveats. The drafting pipeline here is part of the broader Crazystack developer education ecosystem, which also runs the Bootcamp do Dev Doido and Crazystack Typescript programs.
Flows reward iteration: start from a template, hill climb the prompt, and let the eval results decide where the next token goes.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits