back to work

axo-agent

An autonomous agent that tests how AI-ready a website is.

See it working

Problem

AI agents are becoming real visitors: they browse sites, read docs, and try to complete whatever task their user gave them. Most websites have no idea how well they handle that, and the emerging standards for it (llms.txt, agents.json, structured data) are invisible unless you go looking.

axo-agent answers with a real AI, not a checklist. A reasoning agent is dropped onto your site with a goal and figures the rest out on its own: it reads the page, decides where to click, hits dead ends, backtracks, and keeps going until it succeeds or gives up. Finding pricing, docs and contact are the scored tasks today, but nothing is scripted; give the agent any goal and it will hunt for it the way an AI assistant would for its user.

Architecture

Decisions and trade-offs

1. A reasoning agent, not a simulator.

A LangGraph ReAct agent drives a headless browser in a plan, act, observe, re-plan loop, the same loop a human tester runs. Its task performance is 60% of the score, and every decision it makes streams to your screen live.

2. Deterministic checks run in parallel, not in the prompt.

Six standards checks are plain code with fixed scoring, so the same site always gets the same baseline. The LLM never grades what a parser can verify.

3. Everything streams live.

Redis pub/sub pushes every agent step over SSE to the browser, so you watch the agent think instead of staring at a spinner.

What I measured

6 deterministic checks · 40% of score

agent task completion · 60% of score

live event stream over SSE