Claude Just Learned to Run Its Own Team of AI Agents

Your assistant just got promoted. It can plan a job, hire a crew for it, and check the crew's work, on its own. The only thing left for you to decide is what to ask it to build.

Dana's Monday morning problem

Dana runs a four-property boutique hotel group, Lisbon, Porto, Faro, Sintra. Every Monday she opens Google, Booking.com, and TripAdvisor and reads the weekend's reviews by hand. Not because she loves it. Because it is the only way she trusts the summary.

Three hundred reviews a month, across four properties, in three languages. A pattern that matters, breakfast running out by 9am in Faro, say, is buried inside noise: one-off complaints, five-star raves, a guest venting about parking. Dana used to ask an AI tool to “summarise this month's reviews.” It gave her a tidy paragraph. She stopped trusting it the week it invented a complaint about slow Wi-Fi that no guest had actually made.

So she went back to reading them herself. One person, one inbox, one Monday morning, forever.

You are the bottleneck, not the model

Here is the part worth sitting with. The model was never the problem. A single AI chat is genuinely good at reading a review and telling you what it says. What it is bad at is holding three hundred of them in its head at once, and it is worse at telling you, honestly, which of its own conclusions it is sure of.

That second part, an AI checking its own homework before handing it to you, used to require you to do the checking. You read the summary, then you skimmed the raw reviews to see if it was right. You were the person asking the question, the fact-checker, the project manager, and the one waiting between every step, all at once.

That is the part that changed in 2026.

What are dynamic workflows in Claude Code?

Dynamic workflows are a feature Anthropic shipped inside Claude Code: instead of Claude answering you turn by turn, Claude writes a short script, in plain code, that plans a job, splits it into pieces, hands each piece to its own AI subagent, and has a separate agent check every finding before it reaches you. The script runs in the background while your own conversation stays free. Anthropic's own description is blunt about the scale: workflows can run “tens to hundreds of parallel subagents in a single session, checking its work before anything reaches you.”

It is generally available now, on by default for Claude's Max, Team, and Enterprise plans and through the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Pro plan users switch it on inside /config. You do not write the orchestration script yourself. You describe the job in plain English, and Claude writes and runs the script for you.

Who decides what happens next, and why that is the only question that matters

Claude Code actually gives you five different ways to split work across agents, and creator Amir Zaji laid them out cleanly by testing all five on the same real project, building a small website page by page. His organising question cuts through the jargon: for each approach, who decides what the next step is, you, Claude, or a script?

• One session. You decide every step, one prompt at a time. Full control, but you wait between every instruction, and the AI's memory of your brief fades as the conversation grows. In Zaji's test, one simple website build already used 44 percent of a one-million-token context window.
• Parallel sessions. You open two or three separate chats and give each one a task. They run at once, but none of them knows what the others are doing, and all of them report back to you separately.
• Subagents. Claude decides. You give one instruction, Claude spins up several workers, each with its own memory, and only the summary lands in your main chat.
• Agent teams. Like subagents, but the workers can talk to each other directly before reporting back. Still experimental, and only available from the terminal.
• Dynamic workflows. A script decides. Claude writes the plan as code first, then runs it. Because the plan lives outside the conversation, it can send a finding to a second, independent agent for verification before it ever reaches you, and you can save the whole script and rerun it next month.

Anthropic's own documentation makes the same point in one table: for subagents, skills, and agent teams, Claude is still the orchestrator, deciding turn by turn what happens next. For a workflow, the plan is code, so a run can stretch to "dozens to hundreds of agents" instead of a handful, and it can be interrupted and resumed instead of restarting from zero.

Why this is worth paying attention to right now

That is an extreme, expert example, a full systems-language port, not something a hotel group needs. But the shape of it matters: one workflow mapped a technical detail for every part of the code, the next wrote every file as a parallel job with two independent reviewers checking each one, and a fix loop kept running until the tests were clean. Nobody sat there approving each of the hundreds of small steps.

Zaji's own small-scale test on a website audit put a number on the everyday version: a workflow that checked every page against a brief, checked for broken links, and verified mobile display used fourteen agents and about 660,000 tokens, while his own conversation, the one holding the plan, stayed under 75,000. The heavy lifting moved off his plate and onto a crew he never had to manage directly.

Read the two side by side and the pattern is the same one Anthropic engineers have been repeating online this year: the model was rarely the ceiling. The shape of the work you hand it was. One well-known Anthropic-adjacent post making the rounds on X this week put it as “90 percent of our engineers were using self-improving loops. Now everyone shifted to building agentic graphs. No more prompting.” Read literally that overstates it, plenty of good work still happens one prompt at a time, but the direction is real, and it has stopped being a developer-only story.

Isn't this just going to blow through my AI budget?

Fair question, and Anthropic does not hide the answer: dynamic workflows use meaningfully more tokens than a normal conversation, because you are paying for several agents' worth of reading and re-checking instead of one. Their own advice is to start with a small, scoped task before you point a workflow at anything big, and the first time a workflow runs, Claude Code shows you what is about to happen and asks you to confirm it.

The honest way to think about it: you are not paying for one AI reply anymore. You are paying for a small team's worth of work, done in minutes instead of a week of Mondays. For a task that used to cost you actual hours, that trade is usually an easy one. For a five-second question, it is overkill, use a normal chat.

How Dana would actually run this

Here is the version Dana could run without touching a line of code, adapted from the exact kind of instruction that works in practice:

1. Collect the raw material. Export or paste the last 30 days of reviews from Google, Booking.com, and TripAdvisor into one folder, or connect the review source Claude Code can read directly.
2. Describe the job in plain language, and ask for a workflow. Something close to: “Use a workflow to read every review from the last 30 days. Tag each one by category, staff, cleanliness, noise, breakfast, price. Only report a pattern if at least three independent reviews say the same thing. Have a second agent verify each pattern before it is reported, and list anything it cannot confirm as unverified rather than dropping it.”
3. Let it plan before it runs. Claude writes the orchestration script first and shows you the shape of the job, how many agents, roughly how much it will use, before anything executes. You approve it once.
4. Watch it work without babysitting it. The run shows its phases in the background, reading, tagging, cross-checking, while you keep working. Nothing lands in your inbox until it is done.
5. Get one report you can actually act on in a Monday meeting. A short list of verified patterns, each one tied to the reviews that support it, not an AI paragraph you have to double-check yourself.
6. Save it. Ask Claude to save the whole thing as a reusable command, its own documentation calls this a workflow you can rerun by name, so next month is one instruction, not a rebuild.

What you have after this

A week in: one clean, verified report replaces a folder of raw reviews nobody had time to finish reading.

A month in: the workflow is a saved command. Running it again costs one sentence, not a rebuild, and Dana starts trusting the report enough to bring it straight into her Monday team meeting.

Six months in: the same approach is quietly running on more than reviews, supplier invoices checked against delivery notes, a monthly audit of whether every location's menu still matches the brand's actual recipes. None of it required Dana to learn to code. All of it required her to stop asking one AI question at a time and start asking for a job to be planned, run, and checked.

Same reviews. Same four hotels. A completely different Monday morning.