Stack
The AI-first toolchain I actually build with.
Claude Fable 5.1
For the demanding tasks and orchestration: planning, large refactors, reviewing what the subagents ship.
PROS
+Holds a large refactor together
+Best judgement on architecture and API shape
+Orchestrates subagents reliably
CONS
−Overkill for trivial edits
−Cost adds up on long runs
Claude Opus 5.5
The workhorse in workflows and subagents. Takes scoped tasks from Fable and ships them.
PROS
+Fast and cheap enough to fan out
+Follows verify-then-commit loops
+Strong on well-scoped coding tasks
CONS
−Needs a tight spec
−Less taste on architecture calls
GPT-6 Astra
Orchestrator in Codex. My sparring partner for product decisions, and the one I let run overnight once the plan is fixed.
PROS
+Good at thinking through what to build and what not
+Holds up on long unattended runs
+Orchestrates Sol subagents
CONS
−Needs the plan fixed before it runs
−Follows existing patterns in the codebase too readily
GPT-6 Sol
Second worker and second opinion.
PROS
+Fast on isolated tasks
+Strong at test writing
+Cheap to run in parallel
CONS
−Drifts on large context
−Needs tighter specs
All of my code is written by agents. I review, intervene and steer it toward the outcome I want.