efficient-frontier

Efficient frontier

Use the expensive model where judgment matters. Push repeatable, bounded, or token-heavy work to cheaper/faster subagents.

Source: BuilderIO/skills (efficient-frontier). MIT.

For Claude Fable specifically, prefer efficient-fable.

Workflow

1. Frontier-only: architecture, prioritization, ambiguity, risk, synthesis, final review
2. Delegable: repo scans, search, docs extraction, testing passes, log reduction, narrow coding, mechanical edits
3. Spawn parallel subagents with clear ownership, bounded scope, verification gates, expected evidence
4. Require compact returns: findings, files, commands, residual risk, stop conditions, decisions needed
5. Integrate and review centrally before presenting

Handoff packets

Self-contained prompts (assume no chat context): repo path, objective, scope, out-of-scope, targets, return format, verification commands, stop conditions.

Useful stops: code ≠ assumption; verification fails twice; needs files outside scope; no concrete evidence.

Review loop

Treat delegated output as evidence, not verdict. Reopen cited files, skim high-risk diffs, spot-check verification. Resolve disagreements at the frontier layer.

Soft scenarios

• Research — delegate scans/extraction; keep judgment
• Coding — bounded patches when ownership is clear; integrate centrally
• Testing — frontier chooses strategy; cheaper agents run checks/logs and return exact failures
• Debugging — independent theories in parallel; final diagnosis stays frontier

Guardrails

• Do not delegate the immediate blocker you need next
• Do not have multiple agents edit the same files concurrently
• Do not trust high-risk conclusions blindly — inspect evidence
• Pattern works best when exploration/implementation can parallelize

Default framing

Use the frontier model as orchestrator and reviewer; cheaper subagents for token-heavy research, coding, or testing so expensive tokens go to judgment and final quality.