Agents
A chat model answers in text. An agent goes further: it explores your repo, uses tools, iterates when something fails, and stops for your OK on risky steps. You describe the outcome; the agent does the legwork — you review the result like any other contribution.
What this page is for
Foundation order: How AI models work → Tool calling → Agents → Working with agents. For your first run, see Quickstart.
Who this page is for
What an agent is
An agent is not a smarter autocomplete. It is a system that combines:
Together, these let the agent work toward a goal instead of stopping at a single paragraph. The model proposes; tools gather facts; you approve sensitive steps and merge when ready.
That last row is the shift many teams feel: you move from doing every keystroke to defining tasks, reviewing output, and steering when the loop goes off track.
The agent loop
Most agent work follows the same rhythm — whether you are in the desktop app, the CLI, or the cloud:
For a large change, ask for a plan first — outline steps before wide edits. That mirrors how experienced teams use agents: plan → execute in chunks → verify each chunk, not one giant "rewrite everything" message.
When a task is too big for one pass, the agent may delegate a focused sub-task and return a summary to the main conversation — you stay in one thread without every intermediate step cluttering chat.
Where agents run
4RGE exposes the same agent idea on different surfaces. Pick by how interactive you need to be:
Local agents use the folder you opened on your machine. Cloud agents run in an isolated environment your org configures — same review mindset, different place and duration.
Use local when you want to pair closely with the agent. Use cloud when the job is long, parallel, or can wait until you are back at your desk.
You stay in the loop
Agents are powerful because they act — that is also why 4RGE keeps you in charge:
Agents can still be wrong — Hallucination & limitations. Green tests and your judgment beat the agent saying "done."
You can also queue follow-up instructions while a run is busy, or start a fresh conversation when you switch features — see Working with agents for day-to-day habits.
Habits for delegating well
- One outcome per run — "fix failing auth test" beats "clean up the repo."
- Plan before big edits — especially multi-file or cross-service work.
- Say constraints aloud — tests that must pass, files off limits, no push.
- Review like a teammate's PR — diffs, commands, and test output.
- Split when context fills up — Context — or the agent repeats mistakes.
- Match surface to job — local for pairing; cloud for marathon refactors.
- Verify with real signals — CI, staging, or reproduction steps you trust.
Leads: treat agent output as team code — same review bar, same security awareness. Onboarding should cover approvals and scope, not only prompts.