Developer tools · Prototype
Aegis
AI incident response where model confidence is never authorisation.
In plain words. When a website or app breaks, this tool helps engineers work out why. An AI gathers clues and suggests a fix, but it cannot act on its own: fixed safety rules and a human approval have to pass first.
The problem
During an outage the hard part is establishing what happened and which evidence to trust. An agent that guesses a root cause, or acts on a plausible fix, adds operational risk.
What it does
An investigation control plane. Alerts become durable jobs; a 12-node LangGraph workflow gathers evidence across telemetry, topology and code changes, forms hypotheses, and proposes remediation that must clear deterministic gates before anything can run.
Key decisions
- Evidence is append-only with provenance. An unavailable source stays unknown instead of reading as healthy.
- Confidence derives from coverage, corroboration, source reliability and contradictions rather than the model's self-report.
- Executors accept only a validated action type that can be constructed after policy and approval, so prompt text cannot reach an executor directly.
- Postgres is the system of record; Redis is only a cache and event bus.
Limits and status
- Demo-mode measurements say nothing about recovery time for real systems.
- Counts such as node and tool totals are tied to one repository revision.