The announcement
On October 9, 2026, Anthropic put dynamic workflows into public beta on Claude Managed Agents, its service for running AI agents in the cloud. The idea: Claude writes a program that splits the work, launches many agents, passes results between them and merges everything at the end. Meanwhile, you keep talking to the lead agent.
The feature has existed in Claude Code since May 28; now it reaches the managed cloud infrastructure, with its own limits.
1,000 agents, but not at once
Headlines say "1,000 agents in parallel," but the docs are more precise:
- Up to 1,000 agents started in total per run.
- 64 concurrent threads at most, a value that may change and isn't guaranteed by the API.
- A default lifetime of 24 hours per run and up to 10 open runs per session.
- Agents have independent conversation histories but share the files in the same environment. For code tasks, you must clearly separate each one's scope to avoid conflicts.
- The lead agent cannot send new messages to an agent inside a workflow; corrections and re-checks must be built into the program.
To enable it, use the multiagent_20261001 type in the agent definition. No separate API call is needed: the lead agent decides on its own when to start a workflow.
Anthropic's test: 66 of 70 bugs
Anthropic hid 70 bugs in 116,000 lines of code and ran each approach three times:
- A single agent: found 14, 15 and 27 bugs (20% to 39%).
- Dynamic workflow: found 66 in all three runs (94.3%).
That's a big jump, but it's an internal test. Token cost, time, false positives, and whether the same 66 were found each time were not disclosed. And 4 bugs were always missed.
The cost
The feature itself has no extra fee: you pay each agent's tokens plus $0.08 per session-hour (overlapping time isn't double-counted). But Anthropic itself has said multi-agent systems typically use 3 to 10 times more tokens than a single agent. A senior OpenAI engineer even called agent swarms a "massive waste of tokens."
There's a per-session budget cap, shared by all agents. When it's hit, the run pauses, but in-flight requests finish, so the cap can be slightly exceeded. And note: a run marked completed doesn't guarantee every task succeeded; have the workflow list what failed or wasn't analyzed.
What it's good for
- Reviewing large volumes of documents, such as contracts: one agent per document and another checking them all.
- Code audits on large codebases, with well-divided scope.
- Parallel research and triage with cross-checking.
Getting started
- Pick a task that splits into independent parts.
- Set a budget cap from the start.
- Compare against a single agent on the same data: findings, misses, cost and time.
- Require a report of what wasn't covered.
Sources
- Claude docs, multiagent orchestration: https://platform.claude.com/docs/en/managed-agents/multi-agent
- The Decoder (9/10/2026): https://the-decoder.com/anthropics-claude-can-now-orchestrate-up-to-1000-ai-agents-in-parallel-through-dynamic-workflows/
- XenoSpectrum: https://xenospectrum.com/en/claude-managed-agents-dynamic-workflows/
Want to assess whether AI agents at scale make sense for reviews and audits at your company? KGA IT offers a free consultation: kga-it.com.