Resilient
incident
commander
Halo investigates a live product, holds risky fixes behind a human, and keeps going when models rate-limit or tools fail. It degrades instead of dying.
Launch the war roomDetect.Decide.Verify.
Halo pulls live evidence through MCP tools, diagnoses what's actually wrong, and holds any risky action for a human. Once approved, it executes — then re-checks the product to confirm the fix held.
Detect & Diagnose
Pulls live worker status, ingestion, deploys, and failures through Jaguar's MCP tools and works out the real root cause.
Decide with a human
Prepares a recovery action — like a worker restart — but never touches production on its own. Every write waits for operator approval.
Execute & Verify
On approval it runs the action through the live ops path, then re-checks the product and reports whether the fix actually worked.
Built to stay standing when infrastructure doesn't.
Virtual models on AWS Bedrock fail over automatically. A rate-limit on the primary doesn't kill the run — it routes to a backup.
Read and write tools are split across two MCP servers. Anything that writes is isolated and never executes without a person.
Under stress Halo steps down — normal, then a cheaper model with read-only tools, then blackout where it stops and hands off to a human.
Every step leaves a trace — which model resolved, how many spans, which guardrails fired, and exactly which tools the agent called — surfaced right in the war room.
Ready when
things break?
Watch Halo investigate a real incident, hold a fix behind approval, and tell you the truth about whether it actually worked.