Anthropic shipped dynamic workflows for Claude Managed Agents on October 9, and the benchmark they led with is worth paying attention to: a workflow found 66 of 70 planted bugs in a 116,000-line codebase — consistently, across three runs — while a single agent topped out at 27. The raw count is good. The consistency is the real story.
Workflows vs. Subagents: The Distinction That Actually Matters #
Claude Managed Agents already had subagents. A subagent is a delegated specialist — the parent agent hands off a task, the subagent reports back, the parent can react and follow up. That back-and-forth is the point.
Dynamic workflows are different. The lead agent writes a program — an actual multi-phase execution plan — and the runtime dispatches agents to run it in parallel. There is no follow-up loop. The agents execute, return results, the lead agent combines them. It’s batch processing, not conversation.
Anthropic designed this for “work with many pieces, such as reviewing hundreds of documents.” That framing is deliberately narrow. Not every task needs a workflow. But for the right class of problem — large codebases, document batches, broad test coverage — the performance gap is hard to ignore.
What the Benchmark Actually Shows #
Anthropic planted 70 bugs in a 116,000-line codebase, then ran the test three times each way. The single agent found 14, 15, and 27 bugs. The workflow found 66, 66, and 66.
Two things stand out. First, the scale difference is obvious. Second, and less discussed: the single agent’s variance was nearly 2x across runs. The workflow returned the same result every time. For production use — where you need reliable, repeatable quality gates — variance is a problem. The workflow eliminates it.
The caveat: this is Anthropic’s test, on Anthropic’s codebase, with no cost figure attached. Independent benchmarks have not followed yet. Take the numbers seriously, but don’t treat them as universal law until outside teams reproduce them.
Enabling Dynamic Workflows #
Dynamic workflows are in beta, behind the managed-agents-2026-04-01 header. To turn them on, set the agent’s multiagent field. Full setup details are in Anthropic’s workflow runs documentation:
{
"type": "multiagent_20261001",
"workflows": { "type": "enabled" }
}
Under the multiagent_20261001 type, both workflows and subagents are on by default. If you want workflows only, set subagents to disabled. There’s no separate API call to trigger a run — describe the task in user.message, and the agent decides whether a workflow fits. Guide that decision with your system prompt. Monitor execution through workflow_run.* events on the session event stream.
Cost: No Free Lunch at 64 Concurrent Agents #
Pricing follows the standard Managed Agents model: regular token rates plus $0.08 per session-hour for active execution time. Idle time doesn’t accrue charges. There’s no additional per-workflow fee.
That sounds manageable until you do the math on 64 concurrent agents — the documented maximum thread count — all running against a large codebase simultaneously. At Opus 4.6 rates ($5.00/$25.00 per million input/output tokens), a codebase review at scale gets expensive quickly. Haiku 4.5 ($1.00/$5.00) is the correct default for workflow sub-agents unless the task genuinely requires frontier reasoning. A full pricing breakdown is available here. Anthropic doesn’t say this explicitly. It should.
When to Actually Use This #
Not every task belongs in a workflow. Here’s a workable decision framework:
- Use a dynamic workflow when you have a large batch that decomposes naturally into parallel branches — codebase audits, document review pipelines, broad test generation — and you don’t need to inject corrections mid-run.
- Use subagents when you need specialists with their own system prompts and tools, and the parent agent needs to respond to their output before proceeding.
- Use a single agent for anything that fits in a reasonable context window, requires real-time human input, or where cost is a hard constraint.
The community heuristic that’s circulating: if you can sketch a flowchart with parallel branches and data handoffs between stages, a workflow is probably the right tool. If you’re writing a linear prompt, it isn’t.
Beta Realities #
This is still early. The lack of mid-run correction injection — flagged by developers as a gap — means you commit to the workflow structure at the start. There’s also no clear mechanism for handling partial failures gracefully in the current implementation. Anthropic’s event stream tracks progress, but intervention options are limited. Anthropic’s release notes document current limitations as they update.
Dynamic workflows are a meaningful architectural addition to Claude Managed Agents — the benchmark gap between a single agent and a coordinated workflow is too large to dismiss. The production-readiness caveat is real, but for developers already running complex review or processing pipelines, this beta is worth testing now.