# Working with the agent

> Asking for work, pointing at files, steering mid-task, approving actions, reviewing and undoing changes, and keeping the context window under control.

GBS SE Agent · Guides · Source: https://kit.gramproindia.com/se-agent/using-the-agent · Code: https://github.com/gramproindia/agent

## Asking for work

Type what you want and press <kbd>Enter</kbd> (<kbd>Shift</kbd>+<kbd>Enter</kbd>
for a new line). Say what should change and how you will know it worked:

> Add rate limiting to the login endpoint, 5 attempts per minute per IP, and
> make sure the auth tests still pass.

- **The open file is attached automatically**, with any selected code. The chip
  above the input shows what will be sent; click × to leave it out.
- **Point at anything else with `@`.** Type `@` and a few letters to pick a
  file or folder; its contents go with the message, so the agent starts where
  you meant instead of searching. Attachments are capped at twice
  `gbsAgent.maxToolOutputChars` in total.
- **Paste a screenshot** to send it with the message (up to four). It only
  helps with models that accept images. Screenshots are kept in the live chat
  only and are gone after a window reload.
- **Keep typing while it works.** Your message reaches the agent at its next
  step, so you can redirect it without stopping: *"actually, put it in
  middleware instead"*.
- **Stop any time** with <kbd>Esc</kbd> or the stop button. Finished edits
  stay.

## Watching it work

- **Tool rows:** one line per action, such as `Read src/auth.ts`,
  `Search "login"` or `Run npm test`. Click a row for the full result, and a
  file name to open it.
- **Agent cards:** for broad exploration, the agent delegates to subagents that
  search in their own context and report back. The card shows their progress
  and report; their bulky output never enters the main conversation.
- **Live output:** long-running commands stream into their row.
- **Task checklist:** for multi-step work the agent keeps a plan above the
  input and ticks items off. If it stops early, the list is marked
  *not finished*.
- **Thinking:** models that expose their reasoning get a collapsed
  *Thought process* section.

## Approving actions

Shell commands, and edits to files that run code or configure tooling, pause
for your approval. The card shows the full command, or a preview of the edit,
with warnings when a command chains several programs, contacts the network,
redirects output, or contains invisible characters.

- **Run / Allow:** this once.
- **Always allow `npm test`:** for the rest of the session. Not offered for
  risky or sensitive cases.
- **Deny:** the agent is told and tries another way instead of retrying.

The permission chip under the input sets the mode:

| Mode | Edits | Commands |
| --- | --- | --- |
| **Ask first** | ask | ask |
| **Auto-edit** (default) | apply, except sensitive files | ask, except a few read-only ones |
| **Full auto** | apply | run |

Sensitive files (editor, git and CI config, `package.json`, scripts, linter
and build configs, `.mcp.json`) always ask unless the mode is Full auto.
[The permission gate](https://kit.gramproindia.com/se-agent/permissions-and-security) has the full rules.

## Reviewing changes

- Each turn ends with a summary: how long it took, the tokens used, and every
  changed file with `+`/`−` counts.
- **Undo this turn** puts every file back the way it was before the turn:
  edited files are restored and created files are deleted. You see the list
  first, and files you have changed yourself since are flagged before they are
  overwritten. Originals are held in memory, so undo is not available after a
  window reload.
- The diff icon opens VS Code's diff view against the file's state before the
  turn.
- Files open in the editor are edited through the editor buffer, so
  <kbd>Ctrl</kbd>+<kbd>Z</kbd> works as usual.
- `path:line` references in the chat are clickable, and code blocks have a copy
  button.

>
> Changes made by an MCP tool, such as a component installer, do not go through
> the agent's edit tools, so they are not in the turn summary or undo. Use git
> to review those.

## Managing context and cost

The meter next to the send button shows how full the model's context window
is; hover for the session's token totals.

- **`/compact`** summarises the conversation so far to free space. Use it when
  the meter turns orange.
- **`/new`** starts a fresh chat, which is cheaper than compacting when you
  switch to something unrelated.
- **Past chats** are under the chat title. **GBS Agent: Clear Chat History**
  deletes them all.
- Set `gbsAgent.subagentModel` to a smaller model so exploration runs cheaply.

## Giving it project knowledge

- **`AGENTS.md`** (or `CLAUDE.md`) at the project root is read on every
  request. Keep it short: build and test commands, conventions, things that are
  always true.
- **Skills** hold bigger knowledge that only matters sometimes. See
  [Skills](https://kit.gramproindia.com/se-agent/skills).
- **MCP servers** give it tools that answer questions, such as a component
  catalogue. See [MCP servers](https://kit.gramproindia.com/se-agent/mcp-servers).