Simon Willison
★ 9/10 signal
Fable's judgement
agenticmemorytool-usemulti-agent
What happened
Simon Willison shares a practical workflow for Claude Code and Fable. Instead of hardcoding constraints, builders should let the model use its own judgment (e.g., deciding when to run tests). He demonstrates prompting Claude Code to delegate smaller coding tasks to lower-power subagents (Sonnet/Haiku) based on its own judgment to save tokens. Claude Code automatically persisted this instruction as a markdown memory file (~/.claude/projects/.../memory/delegate-coding-to-subagents.md), establishing a hierarchical multi-agent execution loop.
Why it matters
Demonstrates a practical, cost-saving hierarchical multi-agent pattern managed via natural language memory.
The take
This is a masterclass in modern agentic system design. Letting a frontier model act as a high-level router/manager that spawns cheaper subagents for mechanical tasks is the exact design pattern needed to make agentic workflows cost-effective. The fact that Claude Code automatically serializes this into a persistent markdown memory file shows how memory systems are maturing.
Do this
Implement a 'delegate to subagents' system prompt or memory file in your coding agents to route mechanical tasks to cheaper models (like Haiku) while keeping the reasoning model (like Fable/Opus) as the supervisor.
Don't read this site daily. Get it in your inbox.
The daily brief and Sunday deep dive — distilled, scored, and opinionated. For builders only.