Simon Willison: Let the Agent Decide When to Delegate to a Cheaper Subagent

Source: Simon Willison — 2026-07-03

Summary

Drawing on a fireside chat with the Claude Code team, Willison argues developers should let a coding agent use its own judgement to decide when to delegate work to a cheaper, lower-power subagent model, rather than hand-coding prescriptive routing rules. The same principle — trusting agent judgement over prescriptive rules — extends to decisions like when to write tests.

Key Takeaways

• Argues for judgement-based routing (let the agent decide) over rule-based routing (hand-coded "use haiku for X, sonnet for Y" logic).
• Sourced from a Claude Code team fireside chat emphasizing that prescriptive rules often generalize worse than trusting model judgement.
• Extends beyond model routing to other agent decisions, like when testing is actually warranted for a given change.
• A practical harness-design counterpoint to the common advice of prompt-engineering every decision an agent makes.

Discussion

(No questions yet — ask follow-ups via a Claude Code chat session on this repo; answers get appended here.)