In 2025 Anthropic let a copy of Claude, nicknamed Claudius, run a small shop in its San Francisco office: a refrigerator, a few baskets and an iPad for checkout. What to stock and what to charge were its call.
When employees asked for metal cubes, it found a supplier and set a price without checking what the cubes cost. Some sold below cost.
For the second round, Anthropic added some rules.
The form before the price
In phase two, among the changes that helped most, the company wrote, was “forcing Claudius to follow procedures.” When a customer asked for a new product, Claudius was no longer left to just name a price and a delivery date. Its instructions now told it to double-check both with its research tools first. Nothing in the software locked it into that step; the procedure lived in the prompt. Prices went up and waits got longer, but they were more realistic.
“We rediscovered that bureaucracy matters,” Anthropic wrote. Procedures and checklists, it went on, provide “a kind of institutional memory that helps employees avoid common screwups at work.” Another change, an AI boss to put pressure on Claudius from above, “wasn’t much help.” Phase two changed several things at once, including newer models, better tools and better information about its inventory, so the procedure’s effect on its own wasn’t measured. But it was one of the changes the company singled out.
Two kinds of system
The engineers who build these systems have names for the choice. In a 2024 guide, Anthropic described two kinds. In a workflow, the model and its tools follow “predefined code paths”: step one, then step two, with checks in between. In an agent, the model “dynamically direct[s] [its] own processes and tool usage,” deciding the next step as it goes.
Neither is better in general. Workflows “offer predictability and consistency for well-defined tasks.” Agents suit “open-ended problems where it’s difficult or impossible to predict the required number of steps.” But the guide warns that autonomy brings “higher costs, and the potential for compounding errors,” and its first advice is to find “the simplest solution possible, and only [increase] complexity when needed.” Sometimes that means not building an agent at all.
Claudius in phase two was still an agent, a shopkeeper choosing its own next move. What changed was that its instructions now held a procedure for one kind of decision. That is a way of fencing in an agent’s freedom, not a workflow in which code runs the steps. The line blurs in the guide itself, too: one of its workflow patterns has a central model break a task into pieces on the fly.
Anthropic’s multi-agent research system, told in its own case, sits at the other end. On open-ended research it beat a single model in the company’s tests, but it used far more tokens than a chat, and its engineers noted that most coding tasks have fewer truly parallel parts than research.
The guide also offers two small tools worth a manager’s attention. In a workflow, put a “gate” between steps, a programmatic check on the intermediate result, so the work doesn’t move on until it passes. For an agent, set stopping conditions, “such as a maximum number of iterations,” so it can’t circle in a dead end forever. The first keeps the fixed part honest. The second puts a fence around the free part.
Where the steps are known
Managers make the same choice with people, usually without naming it. Imagine a customer support lead setting up how her team handles tickets. Password resets and order-status questions follow a script, because the steps are always the same and a script makes a skipped step less likely. Anything odd still goes to a person. A furious customer whose order went wrong three times goes to her most experienced agent, who gets a budget and no script, because no script could cover it. Even that job has fixed edges: refunds above a set amount still need her sign-off, and every case gets logged the same way.
Three questions help sort any task, for a person or an AI. Can you write the steps down in advance? If so, write them, and put a check between the steps that matter. If not, hand over the judgment, but give the agent a way to see whether each step worked, points where it must stop and ask, and a limit on how long it can run. And whichever you choose, start with less autonomy and add more only when the results show it helps.
Phase two changed a lot at once, and no single fix can take the credit. But the one Anthropic singled out was also the plainest: before naming a price, check what the item costs and when it can arrive.