Anthropic's new beta has Claude write the script that coordinates up to 1,000 agents
Anthropic opened dynamic workflows in Claude Managed Agents on 9 October, letting one run spawn up to 1,000 agents, 64 at a time. The only performance evidence offered is Anthropic's own planted-bug test.
The Engineer · Build desk
Key figures from Anthropic's own planted-bug benchmark: a dynamic workflow found 66 of 70 bugs; a single agent found 14 to 27 of 70 bugs; the test codebase had 116,000 lines.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Planted bugs found by a dynamic workflow, per Anthropic | 66 of 70 bugs | 12 | |
| Planted bugs found by a single agent, at least | 14 of 70 bugs | up to 27, per Anthropic | 12 |
| Test codebase size | 116,000 lines of code | 12 |
What happened
- Claude writes the plan, branching logic and verification passes as a JavaScript script, and a background runtime executes it.
- The feature sits behind one API field: the agent's multiagent value, set to multiagent_20261001.
- The Claude Code version launched May 28, 2026 alongside Opus 4.8 as a research preview and is now generally available.
Why it matters
- decision Teams that hand-configure a coordinator and a roster of subagents now choose between keeping that structure and letting the lead agent generate it.
- constraint The 64-thread ceiling means a full 1,000-agent run needs at least 16 rounds of threads, assuming each agent holds its thread until it finishes.
- exposure A budget or go/no-go sized on 66 of 70 inherits a vendor's benchmark on one codebase with planted bugs.
The case for taking coordination out of the model starts with context. Addy Osmani argued in February 2026 that multi-agent work became a first-class feature because LLMs perform worse as their context expands [17]. The source illustrates it with a codebase-wide migration, where by file 50 the agent is carrying 100K-plus tokens of prior results [18]. That is the author's example, not a measurement.
The design addresses it directly. InfoQ reported, as the source quotes it, that loops, branching, agent-count decisions and verification passes "all live in JavaScript variables, outside Claude's context window" [7]. The model's context receives only the final synthesized answer [6].
Anthropic's planted-bug benchmark is the only performance evidence the source offers [12]. 66 of 70 is 94 percent, and the single-agent range of 14 to 27 is 20 to 39 percent [15]. That is a gap of roughly 2.4 to 4.7 times [16]. It comes from one 116,000-line codebase and Anthropic's own benchmark. The source does not report cost per run, trial count, or how the single-agent baseline was configured.
For the number to transfer, your defects would have to resemble planted ones, and your work would have to split into chunks a subagent can check in isolation. A planted bug is a bug somebody already knows is there.
The source's conclusion is that "the frontier has shifted from single-agent capability to coordination infrastructure" [13]. Its other evidence is a GitHub Trending page full of agent tooling and YouTube videos about personal agent advisor boards [14]. Both measure attention.
Harrison Stall of Anthropic called these flows "the next version of multi-agent orchestration" and framed the goal as applying more test-time compute to hard problems [10]. In our view that puts the weight on the model and the compute budget. The lead agent writes the coordination program [3], so the orchestration is only as good as the model that writes it. A buyer comparing vendors on runtime alone would be comparing the part that is easiest to match.
We'd treat the infrastructure claim as plausible and unproven until someone runs the 66-of-70 test on bugs nobody planted.
What to watch
- Whether Anthropic publishes the method behind the 66-of-70 result, or anyone reproduces it on bugs that were not planted.
- Whether the Managed Agents beta follows the Claude Code version from research preview to general availability, and what per-run cost is disclosed.
- Whether the 64-concurrent-thread limit changes when the beta ends.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence30
- Adoption
- Insufficient
- Hype gap+40
- Incentives
- Insufficient
- Confidence45
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
On October 9, Anthropic opened dynamic workflows in Claude Managed Agents as a public beta.
- [2]
A single workflow run can spawn up to 1,000 agents, with 64 threads running concurrently.
- [3]
A lead agent writes the coordination program that distributes work, collects results, and manages information flow across all the agents.
- [4]
The feature ships behind a single API flag: a developer sets the agent's multiagent field to multiagent_20261001.
- [5]
Claude writes a JavaScript orchestration script covering the plan, branching logic, agent-count decisions and verification passes, and a background runtime executes it.
- [6]
The model's context window only receives the final synthesized answer.
- [7]
InfoQ reported that loops, branching, agent-count decisions and verification passes "all live in JavaScript variables, outside Claude's context window."
- [8]
The Claude Code version of dynamic workflows launched May 28, 2026 alongside Opus 4.8, uses primitives such as agent(), parallel() and pipeline(), started as a research preview and is now generally available.
- [9]
Previously developers had to manually configure a coordinator and a roster of subagents; with dynamic workflows the lead agent plans the work and writes the orchestration program, and optional pre-configured specialist subagents can be passed in, but the coordination logic is generated, not hardcoded.
- [10]
Harrison Stall called these flows "the next version of multi-agent orchestration" and framed the ambition as applying more test-time compute to hard problems.
- [11]
Filling 1,000 agent slots 64 threads at a time takes at least 16 rounds of threads, if each agent holds its thread until it finishes.
- [12]
Anthropic's own planted-bug benchmark claims a dynamic workflow found 66 out of 70 bugs in a 116,000-line codebase, compared to 14-27 for a single agent.
- [13]
The source concludes that "the frontier has shifted from single-agent capability to coordination infrastructure."
- [14]
The source cites GitHub Trending as essentially an agent-tooling showcase and YouTube's extended tier as wall-to-wall personal agent boards of advisors.
- [15]
66 of 70 bugs is about 94 percent; the single-agent range of 14 to 27 of 70 is 20 to 39 percent.
- [16]
The workflow result is roughly 2.4 to 4.7 times the single-agent result in bugs found.
- [17]
Addy Osmani made the case in February 2026 that multi-agent capability became a first-class feature because LLMs perform worse as context expands.
- [18]
In a codebase-wide migration by file 50 the agent is swimming in 100K+ tokens of prior context, and its accuracy on file 51 is measurably worse than on file 1.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- dev.toOrchestration Ate the Moat: Anthropic's Agent Swarm
1 article · October 13, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.