Research

QAI Lab studies the four things an agent needs to be useful on real data, and tests the public code agents are assembled from. What the teams learn becomes limits in the platform rather than a paper.

Research teams:

Retrieval

How an agent finds the right fact in a corpus it did not write, and how often it settles for a plausible one. Feeds the Data layer.

Memory

What an agent should carry between turns and between sessions, and what it should be made to forget. Feeds the Data layer.

Tool use

How agents call software they did not write, and how a wrong call is contained rather than merely logged. Feeds Agentic AI.

Evaluation

How to tell whether an agent was right, at a volume no one can read by hand. Feeds the limits enforced in Cloud and Infra.

What the teams have found

Agents are assembled from public repositories, skills and models that almost nobody audits. We test them, and the failures we find become limits enforced underneath the agent.

300
Models evaluated
5,000+
Security issues found in public agent repos and skills

Open source

Code and datasets are released where releasing them helps someone else avoid a failure we already hit.

Explore on GitHub