Who watches
production
at 3 A.M.?
AEGIS is an autonomous on-call SRE: it watches live telemetry, diagnoses incidents, and repairs them — but it never holds root. Every action must clear a zero-trust policy gate.
Capability
≠ Authority
A system prompt that says "please be careful" is not a safety model. Aegis separates what the agent can propose from what it is allowed to execute — the boundary lives in a Pomerium zero-trust proxy, at the network layer, outside the model's reach.
Seven stages,
three sponsors
52.93%
error rate
We ship checkout-api v1.5.0-rc1 — a bad release. Errors spike while latency and CPU stay low. Aegis reads it straight off the live Nexla feed ~7 seconds after injection and calls it: not load — a bad deploy.
HTTP 403
The agent's first move is the standard playbook: restart_service. Pomerium says no.
24s to recovery
$0.02 USDC
on Base
Last stage: telling the humans. Aegis discovers an email capability on Zero.xyz and pays for it itself — wallet 4.96 → 4.94 — to send its own incident report. A machine paying a machine to deliver the postmortem.
it is a proxy
→ execute → verify → notify
record 1784327295552 · detected ~7s after injection
restart_service is not on the route allowlist
v1.5.0-rc1 → v1.4.2 · recovered in 24s · humans: 0
$0.02 USDC on Base · wallet 4.96 → 4.94