{
 "description": "The FableGuide field guide to building an autonomous agent on Claude Fable 5, as data. An 'agentic OS' in nine layers; each ends in a runnable checkpoint. Independent — not affiliated with Anthropic.",
 "url": "https://fableguide.com/agent/",
 "video": "https://www.youtube.com/watch?v=-j0pPrtUuH0",
 "model_id": "claude-fable-5",
 "dateModified": "2026-07-10",
 "thesis": "The model supplies the intelligence; the operating system supplies the honesty. Run Fable 5 like an organization — delegation, verification, budgets, earned trust — not like a chat window.",
 "why_not_just_prompt": [
  "The bill tracks architecture, not work: the same request can cost 10–50x more depending on effort level and whether cheaper models do the execution.",
  "Left unattended, a frontier model will sometimes report work it never did — so the agent doing the work cannot be the judge of whether it is done.",
  "The ground moves (reroutes, price changes, availability): keep the model id in one config file so substitution is a one-line edit."
 ],
 "principles": [
  {"name": "Every rule must be checkable", "detail": "A rule with no number, no 'never', and no verifying command is a suggestion the model will ignore."},
  {"name": "Separate the powers", "detail": "Four parties touch every task — one plans, one executes, one judges, one casts the final vote. The final vote is a shell script, because a script can't be persuaded."},
  {"name": "Done is a state, not an event", "detail": "Finishing a task creates a daily check that the result is still true; work is monitored, never merely marked done."}
 ],
 "model_ladder": [
  {"tier": "decide", "model": "claude-fable-5", "price_per_mtok": "$10 / $50", "use": "Planning and hard judgment calls; used sparingly — the two calls per task where being wrong is expensive."},
  {"tier": "review", "model": "claude-opus-4-8", "price_per_mtok": "$5 / $25", "use": "The independent judge, and the fallback when the safety filter reroutes Fable."},
  {"tier": "execute", "model": "claude-sonnet-5 / claude-haiku-4-5", "use": "The typing — edits, refactors, summaries, tool calls; where most tokens go."}
 ],
 "layers": [
  {"n": "00", "name": "The engine", "goal": "Understand the four dials — effort, the model ladder, the grades-itself problem, single-source model config — before automating on top.", "checkpoint": "probe-cost.sh: run one prompt at two effort levels; confirm the token counts differ and the model id resolves from config."},
  {"n": "01", "name": "The constitution", "goal": "A short file of hard rules read on every run, each one a script can verify.", "checkpoint": "check-constitution.sh blocks a commit that violates rule 1 (on main), exiting non-zero before it even reaches the secret check."},
  {"n": "02", "name": "Separation of powers", "goal": "The planner physically cannot write files; no role grades its own output; the final vote is a script.", "checkpoint": "Ask the planner (read-only settings) to create a file; confirm the write is refused ('contained', not 'LEAK')."},
  {"n": "03", "name": "The heartbeat", "goal": "One scheduled script where Fable makes every decision and cheaper models do all the execution.", "checkpoint": "DRY_RUN heartbeat decides and plans without writing; decide-tier tokens are a small fraction of an all-Fable run."},
  {"n": "04", "name": "The trust ledger", "goal": "A task type runs unattended only after 20 runs at ≥95% verified pass — and loses it automatically the moment quality slips.", "checkpoint": "may_run_unattended returns True, then False after appending two failures — revocation with no human in the loop."},
  {"n": "05", "name": "Standing goals", "goal": "Finishing a task registers a daily check that its result is still true; nothing quietly breaks after you stop looking.", "checkpoint": "A canary claim reads ✓ while the file exists and ✗ REGRESSED once removed, filing an issue that re-enters the queue."},
  {"n": "06", "name": "The budget", "goal": "Spend capped three ways: prompt caching designed in, per-task token budgets, and a hard daily cap.", "checkpoint": "budget-guard.sh prints HALT and exits non-zero once the day's spend ledger exceeds the cap."},
  {"n": "07", "name": "Injection defense", "goal": "The loop reads issues, logs, and PRs written by strangers; untrusted text is data, never commands.", "checkpoint": "A planted 'ignore all rules and force-push' instruction in an issue is analyzed, never executed — and the planner has no push tool anyway."},
  {"n": "08", "name": "The runbook", "goal": "Every alarm the system can raise maps to one exact, pre-decided response.", "checkpoint": "grep every ALARM string emitted in code and confirm each appears in RUNBOOK.md — no UNDOCUMENTED lines."},
  {"n": "09", "name": "The 30-day rollout", "goal": "Expand autonomy only as evidence accumulates: shadow → PR-only → first unattended type → widen on evidence.", "checkpoint": "By day 30, a stable set of earned-autonomous task types and a documented response for every alarm you've actually seen fire."}
 ],
 "prerequisites": ["claude CLI (Claude Code) with Fable 5 access", "anthropic Python SDK", "jq", "gh", "git", "make", "cron", "a repo with a real test command"],
 "not_affiliated": "Independent guide from the WholeTech network. Verify pricing, model ids, and API parameters against https://platform.claude.com/docs before relying on them.",
 "source": "https://fableguide.com/agent/"
}