Anatomy of a Coding Agent

The model is only one part

A coding agent is more than a language model. The model decides what to do next. The harness (the program around it) runs the loop: it sends the conversation and tool descriptions to the model, executes the tool calls the model requests, appends the observations (file contents, command output, test results) and calls the model again. Tools give hands and eyes: read and search files, edit files, run shell commands and tests, use git, fetch documents, call external services (for example through MCP). Context is everything the model can see at once (instructions, relevant files, history); permissions decide which tools may run without asking. Much of what makes one agent better than another lies in the harness: how well it chooses context, how reliably it edits files, how it recovers from errors, and how safely it handles permissions.

The parts in one picture

Read top to bottom: goal in, verified change out.

goal ("fix failing test") + instructions file + repo map
        |
   [ harness / loop ] <-- tool descriptions, permissions, budgets, logs
        |   ^
        v   | observations (file text, command output, test results)
   [ model ]  --- picks next action ---> [ tools: read, search, edit, run tests, git, MCP ]
        |
   done?  -- yes --> diff + summary for a human to review
          -- no  --> next step (until the step/cost limit)

Read the logs of a run

The sequence of tool calls tells you exactly how an agent reached its answer. Learn to read it.

Quick check: What does the harness do?

• Runs the loop: calls the model, executes its tool requests, feeds back the results
• Trains the model
• Replaces the tests
• Stores the model weights