How to hand your agent a task, give it what it needs, steer it while it works, and recover when a turn goes wrong.

Suzent is an agent, not just a chatbot: it reads files, runs commands, searches the web, and keeps going until the job is done. This page is about working with it well.

Hand it a task

Say what you want done and what "done" looks like. "Summarize these three reports into a one-page brief in brief.md" gets better results than "look at these reports".

One job per chat. A chat keeps its whole history in view, so unrelated work in the same chat only adds noise. Start a new chat for a new job, and group related chats into a project.
Tell it what you know. Mention constraints, preferences, and where things are. What it learns about you is kept in memory, so you don't have to repeat yourself next time.
Bigger jobs can run on their own. For work that takes many steps, set a goal with /goal and the agent keeps working until it's met. For work that should repeat, schedule it. See Automation.

Give it the right material

The agent only sees what you give it.

You want it to use	Do this
A document or image you have	Attach it, or paste the image into the message. It's kept in the project's folder.
A file it can already reach	Type `@` and part of the file name to point at it exactly.
A folder on your computer	Open it with Working Dir for this chat.
Something on the web	Just ask. It searches and reads pages on its own, and can use a browser.

See Workspace & sandbox for what the agent can reach and where its files go.

Pick the model and effort for the job

Each chat has its own model, so you can use a strong model for hard work and a fast, cheap one for quick questions. Thinking sets how long the model reasons first: more effort is slower but more careful, and models that can't reason ignore it.

Background work, such as titling chats and saving memories, uses the models you set in Model roles, not the chat's model.

Decide how much it may do alone

Reading is always allowed. Before the agent changes anything, the chat's permission mode decides whether it asks you:

Ask when you want to see every change, for example in a folder you care about.
Smart for everyday work: low-risk actions run, and a reviewer model checks the rest.
Full Access only for work you'd be happy to let run unattended.

When it does ask, you can allow once, for the session, or always. See Permissions & approvals.

Steer while it works

You don't have to wait for a turn to finish. Type while the agent is working and send it as a Redirect: the agent keeps what it has done so far and continues with your correction. You can redirect or stop a single sub-agent the same way, from its box in the chat.

When a turn goes wrong

What happened	What to do
The answer missed the point	Edit your last message and resend it, or Retry. The turn is rolled back, file edits included, and runs again.
The answer is fine but the file edits aren't	Undo on the "Edited files" card, or type `/undo`. Files go back; the conversation stays.
You want to try another direction and keep this one	Branch from here under any reply starts an independent copy from that point.
The chat has drifted too far	Start a new chat. Memory and the project folder carry over.

Only the most recent turn can be retried or edited. Undo never overwrites a file you've changed yourself since. See Undoing the last turn.

Bring in another coding agent

A chat can be answered by Claude Code, Codex, or another installed coding agent instead of one of your models: choose it under ACP agents in the model picker of a new chat. That agent picks its own model and stays fixed for the chat. See Devices & other agents.

All typed commands are listed in Slash commands.

Working with your agent