Fable's judgement
Simon Willison · Simon Willison · 2026-07-03
Simon Willison shares tips from Anthropic's Claude Code team and community on letting AI coding agents exercise autonomous judgment about testing and model selection, including a Claude Code memory configuration that delegates implementation work to lower-power subagents to reduce top-tier token consumption.
Appears in
Extraction
Topics: claude-codecoding-agentsprompt-engineeringmodel-routingagentic-workflows
Claims
- Giving AI coding agents latitude to use their own judgment — rather than prescriptive rules — produces better outcomes for tasks like deciding when to write tests.
- Delegating implementation tasks to lower-power subagent models (Sonnet for substantive work, Haiku for mechanical edits) significantly reduces consumption of expensive top-tier model tokens.
- Claude Code can persist user workflow preferences as structured memory files in project directories, making them available across sessions.
- Judgment-heavy tasks such as design, auditing, and synthesis should remain in the top-tier model loop, while routine coding can be safely delegated.
Key quotes
Tell Fable to use other models for smaller tasks, applying its own judgement about which model to use.
For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent
So far it seems to be working well. I'm getting a ton of work done and my Fable allowance is shrinking less quickly than before.