« Phase 15 README · Warmup · Deep Dive · Principal Deep Dive · Core Contributor · Staff Notes
Phase 15 — Hitchhiker's Guide
30-second mental model
A coding agent is the ReAct loop with tests as the reward: read the repo → plan → edit via apply-patch (exact search/replace edit blocks) → run the checks → iterate. Spec-Driven Development makes the spec + acceptance checks the contract and the eval. Reusable commands / skills package parameterized workflows. The Temporal/Anthropic framing: use AI to raise quality, not volume — the verification loop is the point.
The facts to tattoo on your arm
| Fact | Why |
|---|---|
| apply-patch = exact search/replace | reviewable, safe; fails on not-found/ambiguous, doesn't corrupt |
| tests are the reward signal | the loop iterates until verify passes |
| spec first (SDD) | the acceptance checks are the contract AND the eval |
| bounded iterations | a loop that can't pass must stop (Phase 00 budget) |
| least-privilege edits | reject patches outside the allowed paths (Phase 09/10) |
| quality, not volume | more code ≠ better; the verify loop is the value |
Framework one-liners
- Claude Code / Codex / Cursor / Aider — coding agents; edit blocks + run tests + iterate.
- SWE-bench — the benchmark for "can an agent fix a real GitHub issue?" (pass@k).
- Spec-Kit (GitHub) — spec-driven development tooling.
- AGENTS.md / CLAUDE.md — repo instructions the agent reads.
- Slash-commands / skills / subagents — reusable, parameterized agent workflows.
War stories
- The agent that "worked" but broke the build. No verify loop; it wrote plausible code that didn't compile. Tests-as-reward is the whole point.
- The patch that corrupted a file. A fuzzy diff applied to the wrong spot. Exact search/replace that fails loudly on ambiguity beats a clever diff.
- The volume trap. A team measured lines-of-AI-code and shipped bugs faster. Quality (review, tests, refactors) is the metric.
Vocabulary
coding agent · apply-patch / edit block · Spec-Driven Development · acceptance checks / verify · reusable command / skill · SWE-bench / pass@k · plan → patch → verify loop · AGENTS.md.
Beginner mistakes
- No verify loop (plausible code that doesn't work).
- Fuzzy diffs that corrupt files instead of failing on ambiguity.
- Unbounded iteration (a loop that can never pass).
- Letting the agent edit anything (no path guardrail).
- Measuring volume (lines) instead of quality (passing tests, clean review).
- Skipping the spec — no acceptance checks means no contract and no eval.