Independent AI-safety research on agent harness design: what changes when a harness grants a model authority over real data, tools, and decisions. Run by Dan Macon. Field notes and articles at alignment.farm.
- Adversarial air gap: a cold white-box adversary found two reproducible breaches in a memory-governance layer; one was closed by a built defense
- Prune to cold store: 57% less hot state at matched answer quality on two engines
- Governed continuity: active-promise recovery without stale revival in a small release fixture
- materialize: fail-closed, hash-audited workspaces from a read manifest
- substrate: a local, turn-based room where humans, agents, and tools share one conversation
Start with construct or the field notes.