Get the Long-Running Agent Operator Playbook
Ten sections, every artifact you need to rebuild Anthropic's long-running agent harness in Claude Code tonight — the same setup Ash Prabaker and Andrew Wilson showed on stage at Code with Claude in San Francisco. Why agents lose the plot on long runs (the 3 root causes: context rot, planning failure, sycophantic self-judgment). The 3-role org chart: Planner, Generator, Evaluator — mapped to PM, IC, QA — with separate context windows, separate system prompts, and the anti-cascade rule that keeps the planner OUT of technical details. The full negotiation contract template — how Generator and Evaluator argue "done" on disk via contract.md BEFORE a line of code is written, with the 27-criteria format that makes critiques actionable. Paste-ready CLAUDE.md harness charter. Paste-ready evaluator sub-agent with the 4-criterion rubric (Design ×3, Originality ×3, Craft ×2, Functionality ×1), the few-shot calibration trick, and Playwright MCP wiring. Paste-ready planner sub-agent with the absolute rules forbidding technical decisions. Paste-ready generator sub-agent with the "cannot mark complete" constraint that stops self-rubber-stamping. The full /longhorizon slash command — every line, drop into .claude/commands/ once. The trace-reading playbook with Ash's "pipe transcripts to files, grep with another agent" debugging trick. And 7 operator mistakes from rebuilding this — including the JSON-vs-MD state file gotcha and the once-per-sprint evaluation cadence that nobody tells you.
Free. No spam. Unsubscribe anytime.