Praxis

Inspiration

Every on-call engineer knows the 3am page for a problem they've already solved. Praxis is an autopilot that takes the first pass — triage, root-cause, and a concrete fix — but stops for explicit human approval before it changes anything.

What it does

Praxis ingests a signed operational alert, uses Qwen models to classify and root-cause it, gathers read-only evidence, and drafts a risk-labelled remediation plan. Nothing executes until a human approves the exact plan. On approval it runs only the approved actions against an isolated target, records every step in an auditable decision trail, and stores a reusable incident memory. It is deployed on Alibaba Cloud Function Compute with a live, public read-only dashboard.

How I built it with Codex and GPT-5.6

I built Praxis end-to-end with OpenAI Codex running GPT-5.6 as my development agent. GPT-5.6's reasoning let me design a strict, fail-closed state machine where no state-changing tool can run before human approval. Codex accelerated the hardest parts: a durable pre-dispatch intent and a terminal "reconciliation required" state, so a real action whose result cannot be verified never reports a false success and is never blindly retried. Codex also wrote and maintained 850+ automated tests and the Function Compute deploy and verify tooling, and every architecture decision is captured as an ADR I designed in Codex sessions with GPT-5.6.

Challenges I ran into

The hardest problem was correctness under uncertainty: a real remediation can succeed while its result-audit write fails. Designing the fail-closed reconciliation model — never invent success, never blindly retry — took careful reasoning with GPT-5.6. Deploying a provisioned, non-idle agent on Function Compute with strict lifecycle bounds and a secret-safe pipeline was the other big one.

What I learned

Human-in-the-loop is not a UI detail — it is a state-machine invariant. And an agent that can take real actions has to be honest about what it cannot verify.