How context is managed in long-running agentic systems
A look at how Director's Journal, Critic's Review and Timeline mechanisms keep multi-agent security investigation systems coherent over long-running sessions.
Maintaining coherence across long-running, multi-agent AI systems poses a significant engineering challenge rooted in the stateless nature of LLM APIs. Since complex security investigations can span hundreds of inference calls, each agent needs a carefully balanced, tailored view of the investigation state — too little context leaves agents disconnected, while too much stifles independent judgment and encourages confirmation bias.
To address this, the team built three complementary context channels: the Director's Journal, a structured working memory supporting six entry types (decisions, observations, findings, questions, actions, hypotheses) each tagged with phase, round, and timestamp; the Critic's Review, which scores expert findings for credibility; and the Critic's Timeline, a consolidated chronological record. This structure lets the Director agent track progress, spot dead ends, and course-correct across an unbounded number of investigation rounds.
Real investigation extracts illustrate the approach: an alert triggered by a kernel module load turned out to be a false positive caused by a developer's package installation, with an overly sensitive detection rule to blame. The Journal entries trace the Director's reasoning from initial suspicion to a confident conclusion, backed by agreement across four expert domains. The Critic's separate role — verifying that expert findings are evidence-backed and not hallucinated — adds a critical layer of reliability.
The architecture offers a concrete, tested reference for engineers building production multi-agent AI systems that must balance context window limits against the need for coherent, aligned reasoning across agents.