During a Fireside Chat at AIE with Cat Wu and Thariq Shihipar from the Claude Code team, participants recommended allowing Fable (and to some degree Opus) to make its own choices about routine work instead of prescribing every action. On 2026-07-03, a user named Simon saved a Claude memory instructing the system to delegate coding tasks to lower-power subagents when appropriate.
What happened and why it matters
- The Claude Code speakers suggested that, rather than explicitly telling the assistant when to run tests or which model to use, it's often better to let the assistant judge when tests or other actions are necessary. For example, instead of ordering "only run automated tests for larger features," allow the assistant to decide whether a change warrants a test.
- Jesse Vincent added a cost-saving tip: ask Fable to use other, lower-capacity models for smaller or routine tasks to conserve Fable tokens.
The saved memory and instruction
Simon prompted Claude Code with an instruction and the assistant stored it as a memory file at:
- ~/.claude/projects/name-of-project/memory/delegate-coding-to-subagents.md
The stored entry includes:
- name: delegate-coding-to-subagents
- description: Simon wants coding tasks delegated to subagents running an appropriately lower-power model
- metadata: node_type: memory, type: feedback, originSessionId: 30068d78-43a9-4fb1-bb29-9799e18c526a
- Stated by Simon on 2026-07-03: "For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent."
The memo explains:
- Why: cost/efficiency — implementation work rarely requires a top-tier model; judgment, review, and synthesis should remain in the main loop.
- How to apply: when a task is primarily code writing/editing, spawn an Agent with a model override (for example, sonnet for substantive implementation, haiku for trivial/mechanical edits) and a self-contained prompt; review the output in the main loop before committing. Design, auditing, data synthesis, and other judgment-heavy work remain with the main model.
Early results
According to the user who saved the memory, the approach appears to be working: they are completing a lot of work while their Fable allowance is shrinking more slowly than before. The setup is intended to let the high-capability model focus on oversight, review, and synthesis while offloading routine implementation to lower-power agents.
Considerations
- The practice aims to optimize cost and efficiency by matching task complexity to model capability.
- It's important to review subagent outputs in the main loop to avoid introducing errors.
Related tags and topics
The saved entry references themes such as claude, ai, claude-code, llms, prompt-engineering, coding-agents, generative-ai, claude-mythos-fable, and anthropic.
