Credit Karma · Incident response
Multi-agent incident debugger
Parallel evidence gathering and grounded synthesis for platform incident investigation.
What this result covers
- Workflow scope
- Platform incident response at Credit Karma, supported by a multi-agent investigation workflow that gathers and synthesizes evidence.
- Measurement context
- An approximate reduction reported in my professional experience. The public summary does not specify the before-and-after periods, baseline MTTR, incident count or severity mix, or the timing boundaries used to calculate MTTR.
Step through the architecture.
Problem
Platform responders had to reconstruct incidents from code changes, documentation, telemetry, cloud signals, and pager activity spread across separate systems.
My contribution
Designed the investigation flow so specialized agents could gather domain evidence in parallel and return it as a grounded root-cause narrative.
Scope and ownership
I owned the investigation architecture from parallel domain evidence gathering through final synthesis. It connects code and documentation, observability, cloud telemetry, and pager activity while keeping the evidence tied to its source.
Architecture decision
I chose specialized agents with a final evidence-synthesis step instead of one generalist agent. The alternative was simpler to operate, but it mixed retrieval and reasoning across unrelated systems. Separate agents kept each investigation grounded in its source, while synthesis could cite and compare their evidence.
ADR 01 · Specialized agents with evidence synthesis
Portfolio decision summary
Alternative considered
A single generalist agent offers a simpler operating model, but mixes retrieval and reasoning across unrelated systems.
Trade-offs & consequences
Separate domain agents keep evidence tied to its source. Final synthesis must preserve provenance and reconcile conflicting signals.
Operating constraints
The workflow needs to preserve evidence provenance, avoid premature conclusions, and make the result useful during time-sensitive incident response.
This case study summarizes my professional work. The diagram is a conceptual overview, and architecture decisions are portfolio summaries.