Thornfield Labs
Currently in Development

The agentic harness for AI safety research.

Most safety research is now done by agents: designing experiments, running evals, reading the results. Thornfield is building the harness that makes that work rigorous, reproducible and safe to run.

Read Our Thesis
01 — Thesis

Safety research has gone agentic. The tooling hasn't.

Researchers now direct agents that write the experiment code, run it across models, and summarize what they found. The bottleneck has moved from doing the work to trusting it.

General-purpose coding harnesses weren't built for this. They blur what an agent did with what it reported, lose the trail from result back to run, and give agents more access than safety work should allow. We think research agents need a harness that treats every action as evidence.

02 — The Harness

Research Orchestration

Plan, branch and run many agent-driven experiments in parallel, with researchers steering at the decisions that matter.

Full Provenance

Every tool call, model query and code change is recorded, so any claim in a write-up traces back to exactly what produced it.

Built-In Containment

Sandboxed model access, scoped permissions and live monitoring, so agents studying risky behavior can't act it out.

Safety-Native Tools

Interpretability, evaluation and red-teaming methods exposed as first-class agent tools, not scripts bolted on afterward.

03 — Status

Heads down, building.

We're working closely with a small group of safety research teams while the harness takes shape. More to share soon.