A self-evolving multi-agent platform.
The lab where I build, test and refine agent patterns on Microsoft Agent Framework, Copilot Studio and Microsoft Foundry, the same stack an enterprise programme runs on.
A supervisor and its specialists.
Specialists with narrow tools under a supervisor that routes the work and logs every decision.
It proposes its own fixes. The evals decide.
EVE reads its own traces and proposes changes to its prompts and tools. Nothing ships until the evals pass.
Nothing ships until the evals pass.
Every behaviour is asserted by an eval before it ships. The same discipline unit tests brought to code, applied to agent behaviour.
Every call traced, every outcome costed.
Cost per agent and per outcome. Whatever an agent does, the trace explains why.
Worked on continuously
New agents and patterns arrive as Microsoft Agent Framework moves.
Orchestration, evals, tracing
The self-evolution loop is research-grade.
Not a public product
I demonstrate EVE directly to teams exploring agents on the Microsoft stack.
You bring the hard question.
A real multi-agent conversation, the evals that assert its behaviour and the cost telemetry behind it.