How AI agents behave on the systems they touch
Aandeg studies what happens when AI agents meet real websites, APIs, and tools, and publishes what we find.
Agents are becoming a new kind of user
AI agents now browse websites, call APIs, and operate tools on people's behalf. The systems they use were built for human visitors or for predictable, hand-written clients. None of them were designed around a model deciding what to do next.
How agents behave in those systems, where they succeed, where they fail, and what they change, is not yet well understood. Aandeg exists to study it carefully and in the open.
From the research
Before the fetch: how AI agents discover what they can do with your site
Long before an agent's harness fetches your page, it already asked two other questions elsewhere - is it allowed here, and what else is available. A tour of robots.txt, Content Signals, llms.txt, .well-known, MCP, A2A, and the newer ARD discovery layer.
Same prompt, different answer: how each harness defines web search and fetch
An agent's web search and web fetch tools are defined by its harness, not the model. Three harnesses, three very different implementations, and why the same prompt can land on your site differently depending on who is asking.
Who's knocking? How AI agents identify themselves to websites
Amazon just blocked Meta's shopping agent for not identifying itself. If you run a website, the same question is already arriving at your server - who is knocking, and should you let them in?
Agent Assay
Agent Assay measures how well an agent can find, understand, and interact with a websites.
Stay in the loop
Get updates on Aandeg and what we're building.