procedural anchoring

Procedural anchoring describes a phenomenon where an AI agent performs better not because it learned some new piece of information, but because it started following a fixed sequence of steps and checks.

Think of a cooking recipe. A recipe usually doesn't teach you new facts about ingredients. Instead, it lays out an order: "boil the water first, then add the ingredients, then flip after 5 minutes." Just following that order noticeably cuts down on the mistakes a first-time cook makes — that's procedural anchoring. It's a similar reason why an agent's performance improves once you give it a task guide: simply pointing out what order to use tools in and what to check along the way reduces mistakes like misconfiguring settings or getting the output format wrong.

But this anchoring has a downside too. If a completely different problem comes up and the agent still forces the old procedure onto it, performance actually drops. The very thing that makes an agent reliably follow a set procedure can eat away at its flexibility when the situation doesn't match.

The article notes that "when agents equipped with a skill performed better, 65.7% of the improvement came from procedural anchoring." In contrast, only 4.5% of the improvement came from the skill directly giving the agent new information it didn't have. It's easy to assume skills help because they inject new knowledge into an agent, but the study's key finding is that defining the order of steps and checks matters far more.

You can compare the same task run with two different prompts.

1. Give only the goal, like "Clean up this file," and run it.
2. This time, add a sequence: "Clean up this file. Steps: 1) First scan the overall structure. 2) Flag duplicate items. 3) After cleaning up, review the result once more."

Comparing the two results reveals which mistakes the agent made not because it lacked new information, but simply because it had no procedure to follow.