AEGIS

Aegis — a governed autonomous on-call SRE agent.

Nobody trusts an AI agent with production root access — so Aegis never has it. The agent watches a live service, diagnoses incidents with an LLM, and remediates fully autonomously, but every action must pass through a zero-trust policy gate it cannot bypass.

How it works: Service telemetry streams through a real Nexla pipeline (webhook source → Nexset), and the agent consumes the feed — citing record IDs in its reasoning. When a bad deploy spiked errors to 52%, Aegis detected it in ~7 seconds and proposed the standard first fix: a restart. Pomerium denied it live (HTTP 403 — not on the route allowlist). The agent adapted, and its rollback was authorized and executed through the zero-trust proxy — recovery in 24 seconds, human actions: zero. It then discovered a pay-per-use Zero.xyz capability and paid $0.02 USDC on Base to email its own incident report.

Zero-dependency Node stack (fake prod service + chaos injection, agent loop, live ops console). Every claim above is from one real logged run — the event log is in the repo.

Built With

• nexla
• pomerium
• zero.xyz