Agents
- Source
- huyenchip.com
- Author
- Chip Huyen
- Date
Context
An 's plan can sound reasonable while calling for tools it does not have or steps that will waste time. In an agent primer, Chip Huyen recommends separating plan generation, validation and execution. A check can reject unavailable actions or an excessively long sequence before the system begins carrying it out.
The plan need only cover the next manageable part of a task. After execution, the system evaluates the result before proceeding. Humans can participate at any of these stages, including approving consequential actions. This makes the boundary between proposing work and performing it an explicit part of the agent's design.
- Agent errors compound multiplicatively. At 95% accuracy per step, a 10-step task lands at 60% and a 100-step task at 0.6%. That arithmetic, not ambition, is why agents need stronger models than single-shot prompting does.
- An agent is pinned down by two things that limit each other: the environment it perceives and acts on, and its tool inventory. A chess agent's only possible actions are legal moves; a robot whose one action is swimming is confined to water.
- Sort tools by whether they read the environment or change it. Read-only tools gather ; write actions let the agent act on the world, which is where both the payoff and the severe failures live.
- More tools is not strictly better. Each added tool makes selection and correct use harder for the model, so the right inventory is found by experiment rather than by handing the agent everything available.
- Web browsing stops a model from going stale, but every internet tool also pipes the open web straight into your agent's context. Choose those APIs deliberately rather than wiring in whatever is easiest.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
“An agent is anything that can perceive its environment and act upon that environment.”
“If the model's accuracy is 95% per step, over 10 steps, the accuracy will drop to 60%, and over 100 steps, the accuracy will be only 0.6%.”
“Just as you shouldn't give an intern the authority to delete your production database, you shouldn't allow an unreliable AI to initiate bank transfers.”
“Personally, I would trust a self-driving car more than the average stranger to give me a lift.”
Checking sign-in…
Loading comments…