0010. Investigation flow: deterministic brief, then a single agent
- Status: Accepted
- Date: 2026-11-02
Context
Most of the value of an AI SRE is in the investigation phase: collecting context faster than a human can. Much of that context is the same for every incident and does not need a model.
Decision
- Trigger. An openstatus, Alertmanager, Grafana or generic webhook, or
stromboli investigate, creates an investigation for a catalog service. Repeats withinSTROMBOLI_DEDUPE_WINDOWare dropped. - Brief, with no LLM. Code collects the service card, dependencies and blast radius, log patterns and metrics against baseline, recent deploys of the service and its dependencies, runbooks and similar past investigations.
stromboli briefprints it on its own, and it must be useful to a human without any model. - One agent with read-only tools then mitigates and investigates (0009), within
STROMBOLI_MAX_TOOL_CALLSandSTROMBOLI_INVESTIGATION_TIMEOUT. - Report. A fixed shape: likely cause, confidence, evidence, ruled out, mitigation, long-term fix, next steps, sent to Slack and the API.
Slack is the primary output: one thread per incident, updated in place rather than spammed, quiet unless there is something new.
Consequences
- The brief is cheap, testable and reusable outside investigations.
- A single agent can anchor on its first guess in long incidents; 0017 proposes separate roles to counter that.