Autonomous coding loops need small stories and fast feedback to work

The Ralph pattern (originally created by Geoff Huntley, popularized by Ryan Carson) — a bash loop that reads a task list, implements, tests, commits, and repeats — shipped 13 user stories in 1 hour. But the loop itself is trivial. The real insight is what makes it work: stories small enough to fit in one context window, explicit acceptance criteria, and test suites that provide fast binary feedback (pass/fail) — a direct application of why Verification is the single highest-leverage practice for agent-assisted coding. But the quality of those test suites matters: Evaluate agent tools with real multi-step tasks, not toy single-call examples shows that toy single-step checks miss the integration flaws that chained, multi-step tasks expose.

Each iteration starts with fresh context (reading updated prd.json and progress.txt), which directly addresses The context window is the fundamental constraint — everything else follows — the agent never accumulates stale context because it resets every cycle. This same principle applies within a single session: Separate research from implementation to preserve context quality argues that even research and implementation should run in separate sessions, because exploration tokens pollute the implementation context. The learnings from each iteration compound via progress.txt, connecting to Compound engineering makes each unit of work improve all future work.

The advanced version (nightly two-part loop) shows the full vision: a compound review job extracts learnings into AGENTS.md, then an auto-compound job implements the top priority using those fresh learnings. You wake up to draft PRs. This is Persistent agent memory preserves institutional knowledge that walks out the door with employees made autonomous — the agent both learns and applies its learnings without human intervention. Karpathy’s Autoresearch adds a sharper framing: Rollback safety nets enable autonomous iteration — not model intelligence — the key enabler isn’t agent intelligence but automatic rollback that makes failures free. And Time-bounded evaluation forces optimization for real-world usefulness instead of idealized performance shows that fixed time budgets per experiment force agents to optimize for real-world throughput rather than idealized quality.