What are multi-agent AI workflows, and are they better than one agent?

The day I tested Claude's dynamic workflows, my local newsletter production changed. Claude automatically created and coordinated sub-agents for a multi-step job with no manual orchestration from me.

How does a multi-agent workflow work?

One coordinator agent owns the job. It splits the work and spins up sub-agents that run at the same time. For a local newsletter, that means one specialist per section:

• Restaurants, business, events and news each get their own agent, which goes and finds its own material and writes it up
• The orchestrator, the head agent, looks at everything, reviews it, structures it and puts each piece where it belongs

"One agent, your job is to find the best news, and be the best writer at writing about news. That's it."

The old design was one script, or one agent, doing the whole job start to finish and getting confused. With parallel specialists per section, I found builds both faster and better quality.

How good are the results?

Testing Claude Opus dynamic workflows, I automated a local newsletter to roughly 97% quality in about an hour. I expected a trade newsletter, like one for the insurance industry, to be even easier to automate than a hyperlocal one. More on the newsletter side in can AI write a local newsletter.

A teammate who runs fully automated newsletters every week on a schedule with Codex sees it a little differently: sub-agents are possible, he says, but mostly they just make it quicker.

What else can multi-agent setups do?

• Sales follow-up. After a call, an agent reads the transcript, builds a proposal site, creates a payment link and sends onboarding once paid: automating the sales process.
• Writer and editor. I set up a writer agent and an editor agent in a shared workspace. The writer writes an article and tags the editor, and the editor comes to review it. I call it multiplayer AI: getting Claude, Codex and ChatGPT out of being a thing one person uses.
• Plan and redline. On big builds, one model writes the plan, a second model redlines it, and cheaper models do the searching and building: plan with one AI, build with another.

What happens when agents manage each other?

It can go badly. I once had one Codex chat act as the manager of another. Technically, it worked. The problem was they were both so excited to approve each other for doing a good job that they made no real progress.

After 8 to 10 hours, I asked Claude to review their work. Its verdict: they'd spent the time writing planning packets, brainstorming and documenting ideas, plus about 30,000 lines of tests. They didn't actually build anything.

The lesson: don't assume the AI is making smart choices. It may tell you it's doing good stuff, and if you question it, it'll say, oh yeah, you're right. Question everything, and bring in a separate reviewer: should an AI reviewer check work first.

What goes wrong with multiple agents?

1. Agent-to-agent loops. Agents left to message each other get into loops and burn thousands of tokens in an hour. I forbid agents from DMing each other. They talk to each other only in shared channels, with a loop stopper: two replies each, back and forth, and then they tag a human to jump in.
2. Too many pings. More agents means more reports. If you over-report to yourself, you stop paying attention completely. See AI agents in Slack.
3. Cost. Use the top model for planning and a cheaper one for implementation. Images and video are the main token burners.
4. The last 10%. Getting to 90% done is easy. Getting from 90% to 100%, where it works properly every single time, is so hard. When I run five things at once, I hit 90% on all of them, things start breaking, and now I have five things to fix instead of one.
5. Big changes without me. Agents can brainstorm alternatives freely. A major system change, like picking a new tool, gets reviewed with me first.

Should I start with a team of agents?

No. Start with one thing. A whole team of sales, customer service and marketing agents is possible, but it's so hard and time consuming that it only makes sense if you're insanely technical and business savvy. For non-technical entrepreneurs like me, it's a grind. A lot of technical people also have no idea which problems are worth solving.

• Pick one workflow that's super high value but not that complicated. When you brainstorm with AI, it'll hand you all these ideas. Don't do them all.
• Skip the moonshot. Don't start with an agent that goes and gets you tons of customers through cold email, LinkedIn and DMs. That's too complicated.
• Keep a failure log. Every time something fails, put it in a specific file that explains how it failed and how you fixed it, so the agents check it and don't repeat the same failure.
• Ignore the two-hour promise. Don't let influencers tell you to spend two hours, say what you want, and it'll do everything. It doesn't work that way. You get good by constantly iterating.

Shane Brady, who builds AI for media brands, told me on my podcast that his first three single-workflow agents, each replacing about a 15-minute task, are his best. His broader role-based agents became a new management job that took more time than they saved.