Explore

How software gets built with AI

Walk one change from a task to live software. Each step adds a stage, and shows what the agent does on its own and where people stay. About 10 minutes.

Step 01 of 06

The spec

Picture a company that sells invoicing software to other businesses, with 200 engineers. A developer there hands a task to a coding agent, an app such as Claude Code, Copilot, Cursor, Codex or Devin that runs an agent's loop of a model and tools. The model, from a lab such as Anthropic, OpenAI or Google, sees only what the coding agent sends it and is paid in tokens. The code lives in the company's repo, with its tests and rules files (instructions for agents, such as AGENTS.md). The task itself is the spec: what to change, and how anyone will know it's done. The coding agent sends it to the model with the rules file and the files its search finds, and the model drafts a plan.

Your guess

The issue says only “Upgrade the PDF library to version 5.” What does the agent do with the gaps?

Make a guess to continue.

The answer is C. It fills the gaps and treats whatever it can check as done, usually “the build and the tests pass”.

What can go wrong now

With a vague spec the agent guesses what done means, and misses what nobody wrote down, like invoices that must still look the same.

How teams handle it

Write completion criteria into the task (Devin's docs ask for them). For anything bigger than a small fix, use a plan gate, where a person approves the plan first: Jules shows its plan, Kiro writes requirements, design and task files, and Claude Code has a read-only plan mode. More in AI coding agents: how an agent's task gets done.

Next:The plan is text. To change code, the agent needs somewhere to build and run it: The agent at work.

Questions people ask

Is the coding agent the same as the model?

No. The coding agent is the app around the model, which builders call the harness: it builds what gets sent, runs the tools and checks permissions. The model only reads what it's sent and writes back text, including requests to use a tool. The same model can sit behind several coding agents, and one coding agent can switch models.

Can text in a GitHub issue steer the agent?

Yes. Issue text, pull request comments and rules files are all text the agent reads, and it may follow instructions hidden in them. That's called prompt injection. Treat anything in an issue or a comment as untrusted input. On GitHub, only someone with write access can start Copilot's cloud agent, and comments from anyone else never reach it.

Does a rules file like AGENTS.md help?

Less than you'd hope. One study found context files gave agents no real gain in finished tasks and raised the cost by about a fifth. I'd keep the file short and written by a person, and put the effort into checks the agent can run: tests, linters and type checks.