Public: incident.io AI SRE

Architectural notes

incident.io's AI SRE is a multi-agent investigation engine that auto-activates
when an incident is declared and investigates in parallel with human responders.

Models: incident.io publicly credits Anthropic Claude (3.5, 3.7, and Claude Code) and
uses OpenAI models (GPT-4o, 4o-mini, o4); cheap or mini models handle classification
and grading while frontier models do reranking and large-context investigation. The
single production model/version is not disclosed.

The agent framework is bespoke Go (a custom ai.RunPrompt abstraction, a prompt tree,
speculative tool calling, 10+ agents across 50+ components) rather than a named
open-source framework.

Retrieval is hybrid: text similarity, LLM summarization, and selective embeddings
stored in Postgres/pgvector, with frontier-model reranking.

Evaluation is a core public theme: evals-as-unit-tests with graders, scorecards, a
prompt leaderboard, YAML eval cases beside Go code, and LLM-as-judge backtesting on
closed incidents.

Observability and integrations are in-house: a custom trace view with cost/latency
tags, integrations to GitHub/GitLab, Slack, Datadog, and Grafana, and an incident.io
MCP for posting back to incident channels.

Not publicly documented: any third-party agent framework or tracing vendor; PagerDuty
and Sentry are not named as agent data sources.