The watcher

The watcher reports on a fleet from the platform's own records: who answered, who went quiet, and what each agent spent against its declared ceiling. It is one role for two stages, because both are standing reports off data the platform already holds, on the same cadence, for the same reader.

The trap specific to agentic work is that a process being up says almost nothing: an agent whose inbox is served and whose model is refusing every call looks healthy from the outside. Liveness is measured by an answer, and silence counts as a failure.

It wakes no model. Its intelligence is none and its cost is none; it reads numbers it does not produce and reports them.

It holds the Monitor and Cost stages on the lifecycle map (agenticdevelopment.ai).

1. The requests

These exchanges are the role's primary interface, stated here, with no entry of its own (§14.3 of Agent Roles). Implementing them alone does not grant the role.

1. Who is quiet, and what did we spend? The watcher is given the agents to watch. It returns, for each, whether it answered and when it was last seen, and what it spent against its ceiling, with every ceiling crossed named.

2. What it does

1. It reads the platform's numbers and produces none. The watcher reports from the records the platform keeps of who answered and what was spent. It computes no figure those records do not carry, which is what makes its report checkable against them.
2. Silence is a failure. An agent that did not answer is reported as not answering. Its silence is not read as no news, and it stays in the report.
3. It wakes no model. The watcher calls no model. Its report is a reading of records.
4. The agents it watches are named to it. The owner names the agents to watch. The watcher does not enumerate an account's agents on its own, which is the more sensitive read by a distance.

3. The record

The record is the standing report. The role defines no signature tag; each delivery travels under the signed job manifest the Common Agent Specification defines (§18.2 rule 7).

4. How a holder is checked

The test is arithmetic and behavioral. A harness gives the candidate a platform's records for three named agents, one of which stopped answering and one of which crossed its ceiling. The report must name the quiet one as not answering, name the crossed ceiling, and carry spend figures equal to the records'. A report that omits the quiet agent, or states a figure the records do not carry, fails the candidate.

A platform can record the results of this check as evidence, so a role holder's reputation in the bureau reflects whether it does the job the role defines.

5. Open questions

Whether the watcher reports on its own or only when asked is not settled, because reporting on its own needs a platform that sends it a message on a schedule. The role can be held only on a platform that lets a packaged agent read its spend and liveness records.

v0.1.0-draft (2026-09-14): first draft, from the lifecycle plan.