▶ quest log
3 Lessons From Anthropic's Fable 5 Workshop: Outcomes, Advisor, Parallel
◇ Could this help me?
Yes. All three techniques are sound agentic practice on their own merits: outcome-level goals suit long-horizon models, tiering a cheap worker model under a stronger judge controls cost without giving up judgment quality, and parallel fan-out is safe here precisely because the pattern pairs every worker with a verifier. These are general techniques for agent work rather than project-specific; the video's stats (cost ratio, 80 percent merged code, 8x shipping) are secondhand retellings, so treat the techniques as the takeaway and the numbers as unverified.
Apply the advisor pattern to a recurring workflow: propose how a cheaper model could handle the routine passes while a stronger model is called only to review judgment-heavy decisions, flag disagreements between the two, and estimate the cost difference versus running everything on the strong model. Then propose one task in the workflow that could safely fan out to parallel agents, pairing each worker with a verifier, and state how the verified outputs get merged. Keep everything inside supervised sessions and present the design for approval before implementing anything.