DIRECTION 2 · AGENTS
Research Agents
Agents that do research the way researchers do — on a gated, verification-first runtime.
Agent reliability is not prompt engineering; it is runtime design. The harness governs a gated cognition loop — define → explore → design the test → plan → gate → execute → verify → crystallize — with spec-driven execution, skill abstraction, layered tests (unit / E2E / domain acceptance) and human-in-the-loop gates.
On top of the runtime, research agents must do more than execute workflows: they chain hypotheses, design the critical experiment, produce control plots, and submit to verification/falsification loops. The lab builds these systems and runs them against real research tasks.
Other research directions