Slack Speeds Up Security Investigations with a Multi-Agent Architecture

Slack's Security Engineering team describes the AI agent system they built to triage security alerts. The project started in late May 2025 as a simple 300-word prompt prototype, but this approach proved inconsistent—sometimes yielding sharp cross-referenced insights, other times jumping to unsupported conclusions. The team found that refining prompts alone couldn't provide the fine-grained control needed for reliable investigations.

Their solution was to decompose the investigation into a chain of separate model invocations, each with a single purpose and a structured JSON output. Drawing on research into meta-prompting (Stanford/OpenAI) and multi-persona collaboration (Microsoft Research), they designed three persona categories: a Director that drives the investigation, domain Expert agents (Access, Cloud, Code, and Threat) that produce findings, and a Critic that scores the credibility of those findings. This weakly adversarial Director-Expert-Critic loop helps curb hallucinations and inconsistent evidence interpretation.

The architecture also implements a 'knowledge pyramid,' letting token-heavy expert analysis run on cheaper models while reserving costlier models for the Critic's synthesis and the Director's high-level reasoning. Investigations are further split into phases—starting with a discovery phase—so the Director can adaptively progress the process. The design offers a concrete blueprint for how security teams can make LLM-driven automation both reliable and scalable.