The most useful thing I have learned about AI coding agents is that they are not junior developers. They are very fast, very literal teammates with no memory of yesterday and no instinct for when a task is a bad idea.

Treat them accordingly.

The work that fits

Agents shine on tasks with a clear definition of done and cheap verification:

  • Mechanical refactors across many files.
  • Test scaffolding for existing behavior.
  • Migrations with a compiler and a test suite as the judge.
  • Investigation: “where does this flag get read?” across a large repository.

They struggle when the task requires taste, stakeholder context, or knowing which of three plausible implementations will still make sense in six months. That is not a prompt engineering problem. It is a specification problem, and it is yours.

Guardrails come first

The teams getting value from agents are not the ones with the cleverest prompts. They are the ones with the boring infrastructure:

  1. Sandboxed execution. Agents run in a disposable container or VM with scoped credentials. No direct access to production, no ambient cloud tokens.
  2. A hard boundary on changes. Every edit goes through a branch, a diff, and a review. Autonomy stops at the pull request.
  3. Fast, reliable checks. Formatting, types, unit tests, and a security scan. If verification takes twenty minutes, the feedback loop collapses.
  4. Small tasks. A task that can be described in two sentences and verified in one command is a good task.

Local models are a real option

Not every workflow needs a frontier model. For codebase search, summarization, and simple edits, a local model served through Ollama or vLLM keeps data on the machine and removes per-token cost from the equation. The trade-off is latency and quality on complex reasoning — so route by task, not by ideology.

Measure the review burden

The metric that matters is not lines generated. It is how much review time each change consumes and how often it has to be redone. An agent that produces a 400-line diff you rewrite by hand has negative value.

Start with a narrow workflow, instrument the review cost, and expand only when the numbers justify it. The goal is leverage, not spectacle.