FAQ

Is this just Ralph?

No — the bare "Ralph" loop (re-run a prompt in while true until done) is the seed, and it's where the lineage starts (see the origin story). Loop engineering is Ralph plus governance: a verifiable Done-when, a separate evaluator (writer ≠ checker), a circuit-breaker, and a cost cap. The bare loop is the engine; the contract and the governance are the brakes and the steering. The worked examples are all governed loops, not bare while true.

Won't it bankrupt me?

Only if you run it ungoverned. A small per-pass cost times thousands of passes is a real bill — that's the whole point of example 1, where the ungoverned run cost ~19× the governed one for the same work ($217.34 vs $11.20, illustrative). The fixes are boring and effective: a Done-when the loop can't argue with, a no-progress halt, a hard cost cap, and long intervals. See the cost math.

Does Cowork (or my tool) support it?

The two primitives — /goal with a separate evaluator and governed interval /loop — are tool-agnostic ideas; most agent tools can express them, with varying durability (see the tool matrix and the durability ladder). The matrix is illustrative — as of June 2026, verify before relying against each tool's current docs. Bet on the primitives, not a specific wrapper.

Is everything here real?

No — it's synthetic and safe by design: fictional orgs, toy repos, made-up datasets, every example labeled "reconstruction for teaching." The receipts (logs, ledgers, .xlsx) are generated and machine-checked to agree with the prose, so the mechanics are real even though the data isn't.

Back to the guide index.