How the agent works

Autonomous loop

Enable the "Autonomous mode" toggle in the chat header and describe the task. The agent:

  1. Plans — analyzes the request and picks the tools it needs.
  2. Acts — calls tools: reads files, edits code, runs commands.
  3. Verifies — inspects each step's result and decides what is next.
  4. Repeats — until the task is done.
  5. Reports — summarizes what was done: files changed, commands executed.

Action timeline

While working, every step is visible: tool name, arguments, result, duration and token usage. Long operations expand to show details.

Plan mode

The "Plan mode" toggle before launching a task: the agent first shows a plan — which files it will read, change, which commands it will run — and waits for your approval. You can reject a plan and request a new one: files are touched only after you press "Approve".

Background tasks

Long tasks can run in the background: the chat stays free, progress is visible in the background task list. The work continues even if you switch to another conversation.

Failure protection

  • Silence guard — if the model returns an empty answer three times in a row, the turn stops with a clear error instead of an invisible empty message.
  • Connection watchdogs — if the provider is silent for 120 seconds mid-response or does not respond within 90 seconds when connecting, the stuck request is aborted: the run never hangs, just send again.
  • Auto-continuation — an answer cut off by the token limit is continued automatically (up to 3 times), marked "Truncated" / "Continuation".
  • Network wait — if the internet drops, the cloud agent waits for recovery (up to 10 minutes) and resumes where it stopped.
  • Context auto-compression — in long runs, older history is folded into a summary (~96K-token threshold): token spend and step latency stay flat late into the task. Originals are kept in the chat archive.
  • Iteration limit — upon reaching the limit, the agent is forced to produce a final report.
  • Three-level patching — exact match → line-based → whitespace-insensitive: edits apply reliably even to slightly changed files.
← Docs