Prompting for Agents | Code w/ Claude
- Source
- youtube.com
- Author
- Anthropic
- Date
Context
Anthropic applied-AI staff define an as a model using tools in a loop, then argue in a Code with Claude talk that agents suit only some tasks. Their checklist asks whether the task is complex enough that the steps are not known in advance, valuable enough to justify the cost, feasible given the tools and information you can supply, and low-cost or easy to detect when wrong. Going from a design document to a pull request fits. A task a person can already lay out step by step probably does not.
For prompting, speaker Jeremy's main advice is to think like the agent: read its tool descriptions and responses and ask whether a person would be confused. Prompts should also state rules of thumb outright. In Anthropic's research feature, the model kept searching after finding an answer until it was told it could stop, and it was given budgets: under five tool calls for simple queries, up to 10 or 15 for complex ones.
- 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.
Checking sign-in…
Loading comments…






