Claude Managed Agents Workflows: 300 Contracts, Cost and Budget

Claude Managed Agents workflows let an agent write a program that runs on Anthropic's servers and hands the work to many agents in parallel. They arrived in beta on October 9, 2026. In the docs example, 300 agents review 300 contracts for under $2 with a small model; with larger models the bill goes up. Here is how a full run works, where everyone gets stuck the first time, and how to protect your budget.

What changed: now the program is in charge, not the agent

Most people will confuse this with the subagents we already had. The news isn't that Claude can split up work, it's who's in charge. Before, the lead agent delegated, read every reply, and everything went through its context. Now the agent writes a program, and that program runs in the background on Anthropic's servers.

In Claude Code, workflows have existed since May 28, but they run inside your session. What Managed Agents adds is that you can leave the work running without your laptop open, while your agent keeps talking with you.

How a run works inside

A run has three layers: the run itself, its named phases, and the threads where each agent works. The program launches agents in parallel, collects what they return, and can repeat a step until a review passes.

The piece almost nobody mentions: each agent's result goes to the program, not to the agent talking to you. That's why its context doesn't fill up with hundreds of readings. And you never see that code unless you ask for it.

The step-by-step walkthrough with the docs example

1. Create the agent with the multiagent field and workflows enabled, all under the Managed Agents beta header.
2. Tell it when to use a run in the system prompt: if asked to review more than a handful of contracts, start one; if it's one or two, read them itself.
3. Open the session and ask, for example: which contracts in this folder have a change-of-control clause? Nobody starts the run for you: the agent decides. You'll see the run-created event with two phases, read the contracts and reconcile findings.
4. Where everyone gets stuck the first time: you interrupt and think it stopped, but it didn't. An interrupt only stops the agent's turn. To stop the run you have to ask it with a message and wait for it to finish.

You know it went well when it answers something like "41 of 300 have the clause", and you also check how many threads failed.

The math: what a run costs

The calculation uses an assumption from the video, not from Anthropic: each contract is 20,000 input tokens and 1,000 output tokens, at published prices.

• Haiku 5.5: each read costs a quarter of a cent, so 300 contracts come to 75 cents. With a second agent reviewing all of them, $1.50. Add 8 cents per session hour and you get $1.58.
• Sonnet 5.5: that double pass comes to about $30.
• Opus 5.5: about $60.

There's a ceiling: 1,000 agents per run. A double pass over 500 contracts already hits it. And the fair comparison isn't against a lawyer, it's model against model.

Three business cases (our own analysis)

Reviewing 300 contracts. A small firm or a compliance freelancer receives 300 supplier contracts before a company sale. One agent per contract looks for change of control, auto-renewal and exclusivity, and a second agent reviews all of them. Out comes a table with clause, page and risk, plus a list of contracts not covered. You charge per package.

Auditing a 400-page site. One agent per page checks titles, links and duplicate content, and another cross-checks the results. Key detail: threads share the same sandbox, so each agent must write to its own file. Also, the server can retry an agent, so every step must tolerate running twice.

Classifying 5,000 tickets. You don't launch 5,000 agents because that exceeds the 1,000 ceiling. Split them into batches of ten: that's 500 agents, with room for retries. First you have to define what a correct label is, or you'll get 500 different opinions.

Who it's NOT for

• If your task is a chain where each step depends on the previous one: there's no independent work to split.
• If it's only a few documents: the docs example says to read one or two without a run.
• If you can't verify the output: more agents mean more things to check.

Doesn't a single agent do the job?

Part of that objection is right. Anthropic wrote in January 2026 that multi-agent systems use three to ten times more tokens and that, outside certain cases, coordination costs outweigh the benefits; their advantage is thoroughness, not speed. But 300 contracts don't fit in one conversation without noise contaminating the reading, and here every thread starts clean.

Rule of thumb: if you can name the independent pieces and the rule for combining them in one sentence, use a run. If not, try a single agent first.

What nobody tells you: "completed" doesn't mean passed

• "1,000 agents" means 1,000 started over the whole life of the run. 64 work at any one time, and the docs don't guarantee that number.
• Retries can push past 1,000 threads.
• A run can finish as completed even if threads failed or one couldn't be created.

To trust the result you need your own pass rule and a coverage report saying which contracts weren't read. All of this comes from the vendor's documentation, and it's in beta.

What breaks in practice

• Budget: it's set when you create the session and can't be added to later. At the cap, runs pause and each thread may overshoot by one request.
• Rate limits: the run's requests count against your API limits, and a thread can stop on a rate error.
• No cheap batch: the Batch discount doesn't apply to these sessions.
• Invisible code: you don't see the code the agent wrote unless you ask.

Claude Code or Managed Agents?

In Claude Code, workflows run in your session and Claude asks for confirmation the first time it runs one. Anthropic says that's how it ported Bun from Zig to Rust: about 750,000 lines and 99.8% of tests passing, in 11 days. That is a vendor figure with no independent verification. In Managed Agents the run lives on the server, doesn't depend on your machine, and you can embed it in your own product.

If the work is yours and one-off, Claude Code. If it's a service others use, Managed Agents.

How to apply it in your business?

Start with 10 pieces, not 300, and compare against a single agent. Here's the checklist to avoid burning your budget:

1. Set the budget when you create the session and leave margin for overshoot.
2. Write in the system prompt when to use a run and when not to.
3. Send the bulk phases to Haiku, with predefined agents, up to 20.
4. Use tools that tolerate running twice, and one output path per agent.
5. Define what counts as passed and demand a coverage report.
6. When it's done, ask it to print the workflow word for word, and save it.
7. Start small and compare against a single agent.

For the short version of this feature, read the full breakdown of dynamic workflows and their limits.

Claude Managed Agents workflows FAQ

When did workflows arrive in Managed Agents?

In beta, on October 9, 2026. In Claude Code they've existed since May 28.

How much does reviewing 300 contracts cost?

Assuming 20,000 input and 1,000 output tokens per contract and a double pass: about $1.58 with Haiku 5.5, $30 with Sonnet 5.5 and $60 with Opus 5.5.

Are those 1,000 agents running in parallel?

No. It's 1,000 agents started over the whole life of the run; the docs indicate 64 working at once, with no guarantee.

How do I stop a run?

Send the agent a message. Interrupting only stops its turn, not the run.

Does completed mean it worked?

No. A run can finish as completed even if threads failed, which is why you need a coverage report.

Conclusion

The news about Claude Managed Agents workflows isn't the thousand agents: it's that now someone has to decide what counts as done. Start small, count the failed threads, and define your own pass rule before handing anything over. Tell me which task in your job has the most independent pieces: contracts, pages, tickets or invoices.

Sources: Claude platform release notes (Oct 9, 2026), workflow runs, multiagent orchestration and pricing.