Interactive guide — 03
Same task, different anatomy
Not structurally. A coding agent is a general loop pointed at a repo with file-editing and shell tools; a computer-use agent is the same loop pointed at a screen with browser and input tools. The 'kind' is set by the tool surface and where the loop runs — which is why the same vendor often ships several.
Pick a task, then switch agent kinds — the capability chips on each step link to the taxonomy.
"The login integration test broke after the session change — find it, fix it, make CI green."
- 01readRuns the failing test, reads the traceback and the surrounding fixtures.Shell access
- 02searchGreps the repo for the session API and every call site that touches it.Repo context
- 03editRewrites the test setup and updates the three call sites to the new signature.File editing
- 04runRe-runs the suite locally; iterates on the one still-red case.Code execution
- 05diffPrints a unified diff for your approval before anything is committed.Git & PRs
Frequently asked
What actually differs between agent kinds?
The loop is the same — perceive, act, observe — but three things differ: the surface it acts on (shell, browser, phone line, inbox), the timescale (interactive session vs unattended hours), and the artifact it produces (a diff, a PR, a resolved ticket, a shortlist). Everything else is packaging.
Is a coding agent a different kind of thing than a general agent?
Not structurally. A coding agent is a general loop pointed at a repo with file-editing and shell tools; a computer-use agent is the same loop pointed at a screen with browser and input tools. The 'kind' is set by the tool surface and where the loop runs — which is why the same vendor often ships several.
Why do the traces share so many steps?
Because they share capabilities. 'Read', 'search', 'edit', 'run' in the CLI trace and 'navigate', 'extract', 'fill' in the browser trace are the same loop wearing different tools — that's what the capability tags on each step mark.
Where do frameworks and protocols fit?
They're not in these traces because they don't perform tasks — they carry the traces. MCP servers, agent frameworks, and agent-to-agent protocols are the pipe; the agents above are what flows through it.