CX AgentDocs Open web app
Agent & tools

How the agent works

Tools, planning, the progress checklist, asking you questions, background work and verification.

2 min readUpdated agenttoolsplantodoverifybackground

When you send a message, CX decides whether it can answer directly or needs tools. Tools are small abilities: search the web, read a page, create a PDF, set an alarm, run a command, click in the browser and about 270 more.

The loop#

  1. Understand – CX reads your message, memory, attached files and the chat so far.
  2. Plan – for multi-step jobs it may write a plan. In Plan Mode (submit_plan) it shows a plan card you can edit: add or describe steps, add a note, then approve.
  3. Act – it calls tools one after another. Each step appears as a small card (e.g. "Search web: …", "Read file", "Shell"). Steps that change something may ask for your approval.
  4. Track – long jobs keep a Plan checklist (todo panel) with done/total and the item "Running now".
  5. Verify – file and code work is checked (syntax checks, tests, ZIP verification) before CX says it is done. A tool card can say "Completed, but not verified".
  6. Answer – a final reply, often with cards or files.

Asking you#

CX asks at most one clarifying question per turn, as a form (choices, multiple choice, ordering or free text). If it needs a password, OTP or API key it uses a secure card instead of the chat, so secrets never go into the conversation.

Limits per turn#

On the Termux settings page (search settings for "Termux"; these limits apply to all tools):

  • Tool steps per turn – how many tool calls one reply may chain (1 to 500).
  • Per-turn wall-clock budget – total tool time per turn (1 to 60 min).

Background work#

If you leave the app mid-answer, a foreground service ("Keeps an AI reply and the tools you asked for running") continues the reply. A notification tells you when it finishes. After a reboot, interrupted autonomous runs are listed in an "Autonomous run recovery" notification.

Token budget#

On long tasks the agent can check its own token use (check_token_usage) against soft/hard caps and slow down or ask before continuing.

Rollback#

File edits made in a turn are listed under AI activity with View diff, Open in editor and Undo all changes when a snapshot was captured.