Managing context in long-run agentic applications

Slack Engineering shares how they manage context in long-running multi-agent security investigation systems. The core challenge is that LLM APIs are stateless, and complex investigations spanning hundreds of inference requests can exhaust context windows. Their solution uses three complementary context channels: a Director's Journal (structured working memory for orchestration), a Critic's Review (credibility-scored findings that filter hallucinations), and a Critic's Timeline (consolidated chronological evidence). Rather than passing message history between agent invocations, these channels serve as online context summarization. The Critic uses a five-level credibility rubric (Trustworthy to Misguided) applied across 170,000 findings, with ~26% falling below the plausibility threshold. The Timeline task further reduces hallucinations by enforcing narrative coherence across all evidence.