Chat and runs

What's a run

A run is one goal executed end to end: the agent reads your request, plans, calls tools (files, web, search, browser, connectors, skills), and reports back. Every step is recorded as it happens, so a run is never a black box: you can expand any turn to see exactly which tools ran, with what inputs, and what came back. Runs survive restarts; reopening one replays it exactly as it happened.

You never need a terminal. Asking "test this and fix what breaks" makes the agent run your project's test command itself, through the same approval gates as everything else.

Starting work

1. Pick a workspace

   The home view asks "What should we build in your workspace?" Selector pills under the composer switch the workspace, branch, and where the run executes.

2. Describe the work

   Write naturally: "summarize my notes", "research these three libraries and recommend one", "fix the failing auth test". For repeatable work, type / and pick a skill; type @ to attach files or connector context.

3. Choose how much freedom the agent gets

   The permission pill in the composer switches between auto (safe actions flow, risky ones ask), manual (everything asks), and plan (read-only). The reasoning pill sets the depth, from low to max; higher levels buy more planning and retries, and max adds evaluator and final-review passes.

4. Send it

   The run starts immediately. For big jobs, Anymo may plan a team; see Swarm mode.

Watching the agent work

Progress is quiet by design. While the agent works you see live one-line entries, "Ran web.extract", "Ran file.write", that update in place. When the turn finishes, they collapse into a single header above the answer: "Worked for 15s". Click it to expand the full timeline: every tool call with its icon, one-line input summary, duration, and status dot. Consecutive tool calls fold into "Used 19 tools" rows so long sessions stay readable.

Your messages sit right-aligned in soft bubbles; the agent's output is plain typeset text. Streaming renders progressively, and older history collapses behind an "N previous messages" row at the top of the thread.

File changes and Undo

When a turn changes files, the thread appends a change summary: a header row with the file count and added and removed line totals, plus an Undo control. Below it, one row per file shows its path and line deltas, expandable to the diff. Undo reverts the turn's changes; because every write went through the workspace layer atomically, reverts are clean.

Any file reference gets an Open affordance listing the editors, file manager, and terminals actually installed on your machine. Set your default in Settings.

Local actions

The chat header has a split button for your workspace's actions: press it to run the default (for example Run or Build), or open the menu to pick another or add a new one. Actions are commands you define per workspace; they execute through the same guarded process path as agent tools, with an interactive approval each time. This is how you, or the agent, run tests, builds, and dev servers without a terminal.

When the agent needs permission

If a step is risky (shell, browser control, anything destructive), the run parks and an approval card appears: what the agent wants to do, the tool, the risk level, and the input. Approve to proceed, deny to skip it; the run continues either way. Details in Approvals.

After the run

• Artifacts. Reports and outputs are saved in your workspace and listed with workspace-relative paths.
• From this run. If the run produced learnings, a quiet panel proposes memories and skills for your review; nothing applies without your approval. See Dreams and Skills.
• History. Every run is searchable. Selecting an old run replays it read-only with a "Viewing history" banner and a summary of what was proposed and approved from it.

Frequently asked questions

Can I stop a run?

Yes. Cancellation is cooperative: pending steps never start, and running steps stop at the next safe point, so the event record stays consistent.

How do I know what it actually did?

Expand the "Worked for Xs" header. The timeline you see is rendered from the run's durable event log, the same record used for audit and replay, not a prettied-up summary.

Does the agent remember things between runs?

Only what you approve. Memory proposals from finished runs wait in Dreams until you accept them.

Which model does it use?

Whichever you configured: a local model server, a router, a proxy, or a hosted API. See Model providers.