Coding agents — when the model touches the repo

Golden rule: An agent is only as good as how fast it notices it is wrong. Do not buy longer prompts. Buy feedback: tests, lint, types, CI, simulators and runnable examples are the agent’s senses.

What it is

A coding agent is an agent loop connected to a repository. The model reads files, proposes changes, runs checks, interprets failures and tries another patch. The difference from autocomplete is real: autocomplete finishes a sentence; a coding agent manages a small work episode.

Closed-loop beats one-shot

The common enterprise failure mode is underinvestment in verification. One-shot codegen feels fast until review becomes the bottleneck. Closed-loop means the agent can run work through automated checks without a human in every step. Humans stay on the boundaries: goal, risk, final merge for high-blast-radius changes.

Speed is not the opposite of quality here. Agent speed is encoded quality: tests, skills, CLAUDE.md / AGENTS.md, MCP tools and consistent codebase conventions. What you cannot run automatically, you cannot scale with background agents or automerge.

Case signal: Spotify × Claude Code

In a conversation between Claude Code creator Boris Cherny and Spotify’s Chief Engineer, the internal Honk system follows the same arc: deterministic migration scripts hit edge-case walls; early LLM one-shot failed; iteration plus a judge layer lifted PR success roughly from the low tens of percent toward ~80 %; stronger models later reduced the need for a separate judge; today the shape is agent SDK in Kubernetes plus real CI (Linux/macOS) and device/UI verification. The lesson travels better than the brand names: measure the loop, then scale autonomy.

Where it works

It works best on tasks with a clear definition of done: a failing test, a specific bug, a small refactor, an endpoint, a UI fix against a screenshot or sim. The easier the result is to verify with a command, the less hell leaks out.

Where it hurts

The agent does not know implicit product decisions, may fix the symptom instead of the cause, and often misses adjacent code paths. It needs tests, diff review, and sometimes a human saying: no, that is a workaround wearing a solution costume. Inconsistent monorepos and missing CI turn agents into confident debt machines.

What to remember

A coding agent amplifies engineering discipline. If the project has tests, small tasks, and clean boundaries, it accelerates it. If the project has unclear boundaries, it just produces technical debt faster. Build the verification loop first; then give the model more rope.