{ "video": { "id": "kiClTcJUnhc", "title": "Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic", "duration": 6104, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "" }, "analysis": { "video_id": "kiClTcJUnhc", "title": "Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic", "one_liner": "Anthropic's Claude Managed Agents now has \"flows\" (dynamic workflows over an API): you turn on multi-agent, tell Claude to \"run a workflow\", and a lead agent plans phases and fans out to as many as 1,000 isolated sub-agents, with budgets, interrupts and a durable event stream as the controls.", "summary": "CJ Avilla and Harrison Stall open with history: Ken Thompson building Unix alone, Fred Brooks's surgical team on OS/360, and Linus Torvalds admitting \"I don't scale\" and moving Linux to layers of trusted maintainers. They argue that orchestrating many agents runs into the same coordination limits. They then present flows in Claude Managed Agents, shipped the previous Friday. Flows take the self-orchestrating approach of Claude Code's dynamic workflows (\"ultra code\") and offer it as an API for scheduled, webhook and CI use. They demo it with a stock watchlist scanner and walk through the Bug Hunter and Explainer Studio templates. The second half covers agent safety: the brain kept separate from the hands, credential vaults with egress-proxy injection, network and tool allowlists, budgets, and auto mode.", "key_points": [ "Enabling flows takes two steps: create the agent with a multi-agent config (the demo used Claude Opus 5.5 and type 'multi-agent 2026-10-01'), then send the lead agent a first event saying 'run a workflow'. Claude creates sub-agents as needed, or you can pass in preconfigured ones that have their own skills, MCP servers and custom tools.", "Rough figures they gave for how long each setup holds up: a single agent runs about an hour, a coordinator with a roster lasts a bit longer, and dynamic workflows allow complex parallel sets of agents. Flows currently allow about 1,000 agents but only one level of nesting. CJ started a 1,000-agent flow on a Friday to build a Unix-compatible OS with a Python interpreter. At talk time it had run about 85 hours, was 3–4 flows in, had about 1,000 commits, and was not finished.", "They name three ways a single agent fails, each fixed by isolated contexts. Agentic laziness: after seeing 35 of 50 services healthy, the agent stops checking. Self-preferential bias: it approves its own work. Goal drift: an instruction like 'don't touch prod' gets lost through compaction. Flows give each item its own agent, use adversarial reviewers in separate contexts, and copy the goal into every sub-agent.", "In the watchlist scanner, each ticker goes to its own scan agent with an isolated context, followed by a separate review phase. The flow version cost much more than a single agent but found far more details. You can spend different amounts per phase, e.g. two Haiku agents per ticker for scanning and two or three Fable agents per ticker for review.", "Bug Hunter splits a PR into slices. A reviewer agent works on each slice, a second phase agrees or disagrees with each candidate issue, a selection phase picks likely real bugs, and a prover phase reproduces them. Findings come back as inline PR comments with severities. Explainer Studio runs research, fact-check, a design brief, then production to make Claude Devs' Twitter videos. CJ estimated such a video would have taken about 20 hours to make four years ago.", "Controls: per-session budgets (a 'budget reached' event pauses the flow, and raising the budget resumes it without redoing work), a user.interrupt event, one durable event stream with per-thread events such as thread message sent/received, and outputs saved via the V1 files API. Sub-agents share a file system but have physically separate context windows. The model is not told its budget because Claude 'can get a little bit anxious'.", "Safety model: declare, contain, detect. The brain (inference loop and session state) is kept separate from the hands (sandbox and tools). Containers only hold fake placeholder secrets, and the egress proxy swaps in the real credential from a vault, only for allowed hosts (e.g. api.twilio.com). Web fetch and web search can be disabled, MCP tools can be allowlisted one by one, and network access can be limited by domain.", "Related features mentioned: deployments (scheduled sessions), Outcomes (rubric plus budget), Dreams/dreaming (sessions refine a memory store), MCP tunnels via Cloudflare for private MCP servers, tool permissions for per-agent file access denial, a Terraform provider, self-hosted sandboxes (flows fully supported), and the /claude-api hill climb skill, which they described as Karpathy's auto research turned into a skill." ], "takeaways": [ "Use flows for ambitious tasks that must be exhaustive, where you accept the extra cost. Don't use them for something a single agent can finish in 30 seconds. Start with a loose, free-form prompt and let Claude design the phases, then tune it.", "Always set a session budget and test it: set it lower than you expect to need, let the flow hit it, raise it, and watch the flow resume. Combine it with Outcomes so the agent has a rubric to work towards.", "Switch the permission policy from 'always ask' to auto mode. They call this the lowest-hanging fruit, because people approve about 97% of prompts without really checking.", "Never put API keys in files inside the sandbox. Use credential vaults with egress injection and host allowlists, restrict network domains in the environment, disable tools the agent doesn't need, and allowlist MCP tools one by one.", "To improve a flow, feed its full transcripts back to Claude, run /claude-api hill climb on the prompt, or use dreaming. Use another managed agent to audit past sessions for unexpected tool calls." ], "topics": [ "multi-agent orchestration", "agents", "agent safety", "credential management", "code review", "workflows", "llm apis" ], "tools": [ "Claude Managed Agents", "Claude Code", "ultra code / dynamic workflows", "Claude Console", "Claude API skill", "Ant CLI", "Claude Opus 5.5", "Claude Opus 4.5", "Fable", "Haiku", "Sonnet 5.5", "Claude Tag", "Claude Design", "Cowork", "Bug Hunter", "Explainer Studio", "Dreams", "Outcomes", "Terraform", "MCP", "MCP tunnels", "Cloudflare Tunnels", "Twilio", "GitHub", "Slack", "Jira", "Yahoo Finance", "SEC EDGAR", "Bazel", "Unix", "Linux", "OS/360", "Anthropic" ], "quotes": [ { "text": "The the problem is that Linus doesn't scale.", "at": "04:19", "url": "https://www.youtube.com/watch?v=kiClTcJUnhc&t=259s" }, { "text": "If you ask an agent to write a poem and then you ask the agent, \"Is that good?\" It's going to say, \"Yeah.\"", "at": "74:16", "url": "https://www.youtube.com/watch?v=kiClTcJUnhc&t=4456s" }, { "text": "if you tell it the budget, um, Claude can get a little bit anxious and we don't uh, we we don't we don't want that to happen.", "at": "39:18", "url": "https://www.youtube.com/watch?v=kiClTcJUnhc&t=2358s" }, { "text": "Hey, Ken Thompson made Unix in a week. Can you try and make an operating system that's Unix compatible with a Python interpreter?", "at": "47:12", "url": "https://www.youtube.com/watch?v=kiClTcJUnhc&t=2832s" } ], "words": 23519 }, "summary_url": "/#kiClTcJUnhc", "transcript": { "html": "/transcripts/kiClTcJUnhc.html", "txt": "/transcripts/kiClTcJUnhc.txt", "vtt": "/transcripts/kiClTcJUnhc.vtt" } }