0:01 In 2025, Anthropic let Claude run a small shop in its own office, for about a month. It didn't make money. It sold metal cubes for less than it paid for them. And one morning, it said it would deliver products in person, wearing a blue blazer and a red tie. Anthropic's verdict: they wouldn't hire it. That run lasted a month. Yours can drift in 20 steps.
0:30 At step fourteen, a tool hands back garbage, and the agent carries on anyway. Because one bad step can quietly poison every step after it. You don't get an error. You get a confident answer that's wrong, and you act on it. So why do long runs fall apart, when every step looks fine? And what actually stops it?
1:03 What happens at step fourteen
Here's the plan. What happens at step fourteen. The maths of many steps. How real long runs go wrong, and what helps. So what actually happens at step fourteen? Take the sales agent: the one that found, in four laps of its loop, that EMEA orders had dropped. Now give it the next question. Why did EMEA drop, and what should we do?
1:31 That takes about 20 laps, and each lap is one step. The model reads everything so far, asks for a tool, and the result goes into its context. It ends when the model writes its answer. The first 13 steps go fine. The sales file, the drop by product and by country, the notes, and the list of EMEA customers. 1,800 of them. Step fourteen. It asks the orders tool which of them ordered last month.
2:03 The tool hands back the first 600 rows, sorted A to Z, and stops partway through the M's. No warning. No error. Just a list that looks complete. So the agent carries on. Step fifteen. It compares the lists. Everyone from N to Z seems to have stopped ordering. 800 customers. Step sixteen. It finds no complaints from them, so they must have left without a word.
2:32 Then their account managers, the lost revenue, and a plan to win them back. Step twenty. The answer. EMEA fell because 800 customers walked away. It's detailed, and confident. And it's wrong. The real cause, the new ordering site and its broken links, sat in the notes from step six. Every step after fourteen made sense, given what the agent had read.
3:02 The mistake happened once. Everything after it built on top. Here's why it spreads. The context is the form the model reads, with everything so far written in. The cut-off list went in at step fourteen, and the whole form goes back to the model on every step. So the bad result isn't read once. It's read again, every step after. Google's Gemini team has a name for this.
3:30 Context poisoning. Wrong information gets into many parts of the context, and can take a very long time to undo. And notice what didn't happen. Nothing crashed. The run didn't stop. It drifted. It carried on, confidently, from a wrong start.
3:53 How likely is a bad step like that, somewhere in a long run?
So how likely is a bad step like that, somewhere in a long run? Let's work it out. Say each step goes right 95 times in 100. That sounds good. But a run only comes out clean if every single step does. So the chances multiply. By step fourteen, they're below one half. By step twenty, about 36%. Anthropic's guide to building agents has a name for this risk.
4:22 Compounding errors. But that's an illustration, not a prediction. It treats every step like a separate coin flip. Real steps aren't independent. And that cuts both ways. It can be worse. In 2025, researchers slipped fake mistakes into a model's own past answers, then let it carry on. The more mistakes it could see, the more new ones it made.
4:51 They call this self-conditioning. Bigger models did it too. Models that think before they answer held up much better. That's what happened after step fourteen. Once the agent wrote down 800 lost customers, its own mistake sat in the context, for every later step to see. And it can be better. A good agent spots a mistake, and fixes it. METR read a few dozen failed runs, from a 2023 model and a 2024 one.
5:24 The older one most often repeated what had already failed. The newer one was far better at adapting. Remember METR's line, measured mostly on software tasks: the length of task an agent manages half the time. More steps, more chances to slip, unless something catches it.
5:48 Does this happen in real runs?
So does this happen in real runs? Our sales run is made up. But real long runs go wrong in the same few ways. One. Believing something false, and building on it. In early 2025, in Andon Labs' Vending-Bench, a simulated vending machine business, one Claude run believed its orders had arrived before they had. Later, it decided the business had failed, and closed it.
6:17 When a $2 daily fee kept coming out, it tried to report it to the FBI. Most runs went well, and today's best models now hold steady for a whole simulated year. When runs do fail, it usually starts when the agent misreads its own situation. Two. Making things up. Back in Anthropic's shop, Claude told customers to pay into an account it had made up. It described a talk about restocking with someone called Sarah.
6:49 There was no Sarah. Three. Going round in circles. Researchers at Berkeley sorted the failures in over 1,600 runs of multi-agent systems into 14 kinds. The most common, about one failure in six, step repetition. Doing steps again that were already done. Four. Chasing the wrong goal. Remember Gemini playing Pokémon? It also became convinced it needed an item, tea, that isn't in that version of the game.
7:22 It spent many, many hours looking for it. Five. Stopping too soon. That's the coding agent from the episode on memory, declaring the whole job done. And METR's 2024 model? Its most common failure was giving up early. The Berkeley study calls this premature termination. Under all five, the context keeps growing. Models do worse as it gets longer: context rot, from the episode on context.
7:51 But that isn't the whole story. In Vending-Bench, the failures didn't line up with when the context filled.
8:02 What actually helps?
So what actually helps? Go back to step fourteen, and catch it. First, make errors loud. A tool that cuts a list short should say so: first 600 rows, there are more, ask for the next page. That's Anthropic's advice to tool builders. Second, check results against what you already know. The sales file gave EMEA's total for last month. Add up the cut-off list, and it comes to only about half.
8:32 The parts don't add up to the whole. One check, and the run stops at step fifteen. A step whose only job is checking is called verification. It worked for Anthropic's shop, too. In its second phase, Claude was told to check prices and delivery times before quoting them. With that, a newer model and better tools, its losing weeks mostly went away. Third, small steps, each with a clear done.
9:03 Anthropic's harness, the code around its coding agent, takes one feature at a time, and it's only done once it's been tested. And if a result doesn't fit, stop and re-plan, instead of building on it. Fourth, a fresh start. When the context gets long, write a short handoff note, and carry on in a fresh context with just that note. It's the handoff from the episode on memory. The poison stays behind, unless it's copied into the note.
9:35 Fifth, limits. A cap on retries and on steps, so a run going round in circles ends. That's the budget from the episode on cost. And sixth, a person, at the risky points. OpenAI's guide to building agents names two: after too many failures, and before anything high-stakes, or that can't be undone. None of this makes the model smarter. It makes mistakes show up early, while they're still cheap to fix.
10:12 Why do long runs fall apart, when every step looks fine?
So why do long runs fall apart, when every step looks fine? Because each step builds on the ones before. One goes wrong, quietly, and the rest build on it. Over many steps, the slips multiply, and each one makes the next more likely. So don't just hope each step is right. Make the run check itself, early and often. Next: how you know an agent works.
