Incident management · Production

Slack Engineering: Managing Context in Long-Run Multi-Agent Security Investigations

The problem

Complex, long-running multi-agent security investigations overwhelm language model context windows and make coherent multi-agent reasoning difficult. Each agent needs a tailored view of investigation state — if agents are not anchored to the wider team, investigations become disconnected and incoherent, but sharing too much information stifles creativity and encourages confirmation bias.

Workflow diagram · grounded in source
1
Security alert triggers investigation
trigger
“teams of AI agents collaboratively investigate security alerts”
2
Director orchestrates via Journal
ai_action
“The Director is responsible for orchestrating the investigation: deciding what questions to ask, which Experts to engage, and when to conclude the investigation”
3
Expert agents gather evidence
ai_action
“Each Expert has a subject domain and tools to allow them to interrogate relevant data sources. At the end of their run, the Experts produce findings, citing investigation artifacts (tool calls) to support their conclusions.”
4
Critic reviews findings for credibility
validation
“The Critic's role is to assess the Experts' work, checking that reported findings are supported by evidence and that interpretations are sound”
5
Critic assembles consolidated Timeline
output
“It is challenged to construct the most plausible consolidated timeline from three sources: - The most recent Review - The previous Critic's Timeline - The Director's Journal”
6
Director concludes or continues
routing
“investigations continue until concluded by the Director agent”
Reported outcome

Three complementary context channels — the Director's Journal, Critic's Review, and Critic's Timeline — maintain coherence across investigation rounds while preserving specialized agent roles, enabling more thorough and trustworthy security investigations than any single agent could produce alone.

Reported metrics
Total findings reviewed170,000
findings classified Trustworthy (score 0.9–1.0)37.7%
Findings below plausibility thresholdslightly over a quarter
specimen Timeline confidence score0.83
Reported stack
get_tool_callget_tool_resultget_toolset_infolist_toolsetsget_tool_info
Source
https://slack.engineering/managing-context-in-long-run-agentic-applications/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Three complementary context channels — the Director's Journal, Critic's Review, and Critic's Timeline — maintain coherence across investigation rounds while preserving specialized agent roles, enabling more thorough a…

What tools did this team use?

get_tool_call, get_tool_result, get_toolset_info, list_toolsets, get_tool_info.

What results were reported?

Total findings reviewed: 170,000; findings classified Trustworthy (score 0.9–1.0): 37.7%; Findings below plausibility threshold: slightly over a quarter; specimen Timeline confidence score: 0.83 (source-reported, not independently verified).

How is this incident management AI workflow structured?

Security alert triggers investigation → Director orchestrates via Journal → Expert agents gather evidence → Critic reviews findings for credibility → Critic assembles consolidated Timeline → Director concludes or continues.

Related incident management cases