How Slack Manages Context in Long-running Multi-agent Systems

"To manage the complexity of long-running agent systems, Slack engineers adopted a strategy involving three context channels: a director's journal, a critic's review, and a critic's timeline. This approach allows for structured memory and validation, essential for maintaining coherence as message history expands."

"The coordinator/dispatcher multi-agent design at Slack enables a central coordinator to make decisions and dispatch requests to specialized agents. Critics play a vital role in evaluating the work of experts, ensuring that findings are credible and corroborated by multiple sources."

To enhance productivity in long-running agent systems, Slack engineers transitioned from chat log accumulation to structured memory and validation techniques. This shift is crucial for maintaining coherence as message history grows, which can overwhelm an agent's context window. Slack employs a multi-agent framework with three context channels: a director's journal for structured memory, a critic's review for annotated findings, and a critic's timeline for chronological data. This system allows for effective evaluation and scoring of findings to ensure accuracy and reliability.

Read at InfoQ