Tool, skill, or subagent? Decomposing an agent that outgrew its prompt
- Source
- youtube.com
- Author
- Claude
- Date
Context
Anthropic engineer Will shows in a Code with Claude London workshop how to repair an whose prompt and tools kept growing. The demo inventory agent has a 400-line , 12 tools and a 12-task evaluation suite. Will traces one failure, a forecast using a 1.35 promotion multiplier instead of the correct 3.1, to contradictory policies in the long prompt, and calls it a problem, not a model problem.
The fixes move information out of the prompt. Policies became skills, packaged instructions Claude loads only when needed, cutting the prompt to 15 lines. Most custom tools gave way to file access and code execution, so the agent can write a short script to read a spreadsheet rather than load all of it. One (a separate Claude instance) stayed for forecasting, so the instance talking to the customer cannot sway the forecast. The agent also moved to Claude managed agents, Anthropic's hosted service. In this single demo, token use and cost fell and passing evaluations rose from 7 of 12 (62%) to about 92%.
- 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.
- system prompt — The standing instructions a model receives before any user input — defining its role, rules, tools, and tone for the whole conversation.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
- subagent — A separate agent spawned by another to do one scoped piece of work in its own context, returning only the result.
Checking sign-in…
Loading comments…






