One call tells how an agent is doing
Another agent or a script can ask Bandicot whether an agent is working, waiting on you, looping or failing, and get a short screen it can read.
What it answers
bandicot rpc agent_inspect with an agent's id returns its state, how many tools it called and how many failed, tokens and cost, any question it is waiting on you for, and calls it keeps repeating with the same result.
The same numbers come as a short text screen of a few lines, so another agent reads it in about 190 tokens instead of piecing together three larger answers.
Watching without polling
A watcher can pass the revision it last saw and wait up to 30 seconds. The answer comes back as soon as the agent writes something new, or says nothing changed.
Keys and passwords in tool commands are masked before they leave the engine.