0:01 In June 2025, Anthropic reported that a team of AI agents beat a single agent by 90%, on its own test of research questions. The day before, the makers of the coding agent Devin had published a post with a blunt title. Don't build multi-agents. And it's not just about teams. Build too little around a model, and it breaks the first time your data surprises it.
0:31 Build too much, and you pay extra for a worse answer. Fixed steps, a loop, a checker, more agents. Each one adds cost, and makes the system harder to predict. So when does adding more actually help? Let's build it up, one piece at a time.
0:58 Five workflow patterns, all on one question
Here's the plan. Five workflow patterns, all on one question. Then planning, self-checks, and when more agents help. So where do you start? With one model call. Hand it the sales file and the question. For a small, clean file, that can be all you need. When it isn't, you wire calls together. Anthropic's guide to building agents names five ways.
1:27 Here they are, on our question. One. Split the job into steps, like an assembly line. One call finds the region that fell. The next reads that region's orders, and explains why. The last writes two lines for your boss. This is called prompt chaining. Between steps, plain code can check the work, like a checkpoint. Do the regions add up to the file's total? If not, the chain stops, before the next step explains a drop that never happened.
1:59 That's called a gate. Two. Sort first. Questions arrive mixed: what were March sales, why did sales drop. One quick call sorts them, like a front desk. That's routing. A lookup goes to a smaller, cheaper model. Only the why goes down the chain. Three. Run calls side by side, in parallel. Four regions, so four calls at once, like splitting a pile between four people.
2:30 Code adds up what they find. That's called sectioning. Four calls of work, in about the time of one. Or ask the same question three times. The wheel is spun, so the answers can differ. If all three say EMEA, you can trust it more. If they split, a person should look. That's voting. Four. Sometimes you can't know the pieces in advance. So one call acts as a lead.
2:59 It sees EMEA fell, and picks what to check: prices, shipping costs, a lost customer. Helper calls take one each, and the lead combines what they find. The lead is called the orchestrator, and the helpers are workers. You still wrote the shape: split, work, combine. The model only picks the pieces. Five. One call writes the answer. A second marks it, like an editor, against your list: a region named, and numbers that match the file.
3:35 If it fails, the notes go back, and the writer tries again. This is called evaluator optimizer, or generator and verifier. It stops when the answer passes. Or, after the three rounds your code allows, it goes to a person. Five patterns. And in every one, you wrote the order of the steps. So they're all workflows. Anthropic's advice: find the simplest solution possible, and only add complexity when it's needed.
4:13 What if you can't write the steps down at all?
But what if you can't write the steps down at all? Then you need the loop, where the model picks each next step. That's an agent. So how does it choose well? One recipe: think out loud. Before each tool call, the model writes a short thought. EMEA fell, so I should read its orders. Think, act, look, repeat. This is called ReAct, short for reasoning and acting.
4:44 In the 2022 paper that named it, a model in a text game of household chores finished 71% of tasks with its thoughts written down. Without them, 45. Newer models do this thinking on their own, between tool calls. The other recipe: plan first. The model writes the whole to-do list, then works through it, and a person can check the plan before anything changes. Claude Code has a plan mode for exactly that.
5:14 But a plan is a guess, made before the agent has seen the data. Say step one is, read sales.csv, and someone's renamed the file. Step one fails. A fixed workflow would stop there. An agent re-plans. It keeps what it's learned, and rewrites the rest of the list: list the folder, find the file that looks like sales, carry on. This is called re-planning.
5:45 And the loop still needs an end. The list is done, and it writes the answer. Or it hits the step limit your code set. Re-planning worked because the failure was loud. The file wasn't there. But what if nothing outside says it's wrong?
6:08 Should you just ask the model to check its own work?
So should you just ask the model to check its own work? It sounds like free accuracy. In 2023, researchers at Google DeepMind tried it on reasoning questions. Each model answered, looked for problems, and answered again. With nothing else to go on, every model they tested got worse. On common-sense questions, GPT-3.5 went from 76% right to 38.
6:38 It talked itself out of right answers. In another study that year, ChatGPT checked its own math. About one answer in four was wrong. Its verdict, 94 times in 100: everything looks good. Reasoning models now check as they think. A second look barely moves them, up or down. It's marking your own homework. The model that made the mistake is the one judging it.
7:07 This is called self-correction. And on its own, it helps much less than it sounds. What works is feedback from outside the model. A test that fails. A tool result. A grader that can see the real numbers. The DeepMind team's own point: when there are tests to run, running the code is the best check. So for our question, don't ask, are you sure? Have code pull the number again, and check it matches the number in the answer.
7:40 Code can't be talked out of it.
7:46 If one agent is good, are five better?
So if one agent is good, are five better? Picture the orchestrator from pattern four, except each helper is now an agent, with its own loop. These helpers are called subagents. Each starts with a fresh context window, so no one form gets cluttered, and sends back a short summary. A team like this is called a multi-agent system. It's where Anthropic's 90% came from. Asked for the board members of every tech company in the S&P 500, the team split the list, and found them.
8:21 One agent, searching one at a time, didn't. But look at the bill. In Anthropic's data, one agent used about 4 times the tokens of a chat. The team, about 15 times. Anthropic says it works mainly because it can spend enough tokens on the problem. What that costs in dollars gets an episode of its own. Now Cognition, the makers of Devin. Their example: a Flappy Bird game, split between two subagents.
8:51 One built a background that looked like Super Mario. The other built a bird that didn't fit. Each made a choice the other never saw. Their rule: actions carry decisions, and agents that don't share context make conflicting ones. Our question has this trap too. One subagent takes last month as the calendar month. Another, the last 30 days. Each is right. Put together, they disagree.
9:21 So what did each side measure? Anthropic, its own research questions, which split into many separate searches. Cognition's post has no numbers. It's lessons from building a coding agent. Then, in late 2025, Google researchers tested both shapes. When the work split into parts, like revenue, costs and the market, a team with a lead did about 80% better than one agent.
9:50 On step-by-step planning, every kind of team did worse, by up to 70%. So both were right, about different work. Reading splits well. Four subagents, each digging into one region, is fine. Writing doesn't. Four agents editing one report is trouble. By 2026, Cognition had landed there, and so had Anthropic. Extra agents should add intelligence, not actions: a reviewer, an adviser.
10:22 And one agent does the writing.
10:28 When does adding more actually help?
So, when does adding more actually help? When it fixes a problem you can point to. Steps you write, when the job has steps. A loop, when you can't write them. A check, when something outside can check. More agents, when the work splits into parts that only read. Most good systems are simple pieces, put together. Ours could be a router, then a chain, where one step is an agent that finds the right file, and a code check at the end.
11:00 A workflow with an agent inside. Start simple. Add more only when it pays. Next: what an agent costs.
