How Slack Keeps a Team of AI Agents From Losing the Plot

Key Takeaways

• Slack uses three specialized LLM agents: Director, Experts, Critic
• Director maintains a structured journal to share context across agents
• Critic validates expert output, preventing confirmation bias and drift
• Separate context channels avoid exceeding LLM token limits
• Approach showcases practical context engineering for enterprise AI workflows

Pulse Analysis

Slack’s multi‑agent security platform tackles a core challenge of large‑language‑model deployments: how to keep dozens of inference calls coherent without blowing the context window. By assigning distinct roles—an orchestrating Director, domain‑specific Experts, and a skeptical Critic—the system creates natural boundaries that limit the amount of text each model must process. The Director’s journal acts as a shared working memory, capturing decisions, observations, and hypotheses in a schema‑driven format that other agents can reference without re‑reading the entire alert history. This disciplined approach reduces token consumption, cuts costs, and prevents the confirmation‑bias drift that plagues monolithic prompt designs.

The Critic’s review channel adds a layer of quality control, flagging inconsistencies and ensuring that expert outputs converge on a reliable conclusion. By separating validation from evidence gathering, Slack mitigates the risk of echo chambers where agents reinforce each other’s errors. This mirrors best practices in human investigative teams, where an independent reviewer challenges assumptions before final verdicts. The result is a transparent investigative trail that can be audited, a crucial feature for security operations that must meet compliance and reporting standards.

Beyond Slack’s internal use case, the architecture offers a blueprint for any organization deploying LLMs at scale. Context engineering—splitting information streams into purpose‑built channels—enables teams to build complex, collaborative AI workflows without sacrificing performance or interpretability. As enterprises increasingly rely on AI for threat detection, incident response, and operational analytics, adopting Slack’s modular, memory‑aware design can accelerate deployment, lower inference costs, and improve decision quality across the board.