The harness outlives the model.
We route work to whatever can do it cleanly: large models for hard reasoning, smaller models for repetitive passes, scripts for things machines should not be philosophizing about.
Method
We don’t hire models. We build harnesses.
belt.works is an AI-native shop: small machines, specialized agents, local repos, real checks, and human judgment at the seams. The model is not the product. The harness is.
model agnostic smallest useful model tooling over theater diffs or it didn’t happen
Use the smallest thing that works.
Expensive intelligence is saved for expensive uncertainty. Most software work is reading, shaping, testing, and checking. Paying a giant model to count files is how civilization ends: slowly, then on an invoice.
The output has to survive contact with git.
Agent sessions produce branches, diffs, commits, checks, and reviewable handoffs. If it cannot be inspected, tested, or rolled back, it is not finished.
Mesh roles
Operators, not mascots.
Janine routes and coordinates. Peter handles implementation passes. Egon plans, reviews, and does technical cleanup. The point is not pretending agents are magic employees. The point is giving each worker a narrow lane and a hard stop.
Input: messy goal, stale context, existing code
Harness: scoped prompts, local tools, git branches, tests
Workers: model-specific sessions with narrow responsibilities
Output: pull requests, deployable changes, documented decisions
Rule: no demo theater where a checklist would do.
Operating rule
One money project. One proof project. One infrastructure project.
Everything else waits unless it earns its way back.
Money: client work / intake
Proof: belt.works / public evidence
Infrastructure: mesh status / tooling
Rule: if it does not support survival, evidence, or method, it is drift.