Harness + Forge Integration
Relationship
The harness is the runtime that executes skill graphs. Forge (forge-v2) is a subsystem within the harness that authors new skills: given a text brief, forge emits a complete native skill package (graph.json, modules, scripts, tests).
Forge is itself a harness skill — its topology is declared in a graph and executed by the harness runtime. This means forge benefits from the same deterministic routing, verification batteries, and atomic emit that production skills use.
Forge profiles
| Profile | Topology | Use case |
|---|---|---|
power |
GoT-basic, best-of-3 | Production skills; highest quality |
standard |
GoT-basic | General-purpose; balanced |
minimal |
Linear | Quick prototypes |
The forge cycle
- Brief — the user provides a text description of the desired skill.
- Design — forge decomposes the brief into nodes, edges, topology, and constraints.
- Emit — forge writes the skill package atomically.
- Verify — forge runs the verification battery (graph validation, wiring check, smoke tests).
- Evaluate — forge scores the emitted skill against acceptance criteria.
Harness-forge co-evolution
The harness and forge co-evolve: improvements to the harness (new node types, edge types, verification checks) automatically benefit all forge-emitted skills, and forge improvements (better topology design, tighter acceptance criteria) raise the quality of all future skills.