The Harness
A CEO-model orchestrating disposable frontier-model workers — each isolated in its own git worktree, dispatched through a zero-markup model gateway. Same org, now genuinely multi-model.
Layers
| Layer | Tool | Why |
|---|---|---|
| Orchestrator | Claude Code (Fable) | Plans, dispatches, reviews, merges |
| Dispatcher | tools/worker.py | Worktree lifecycle, logs, cost |
| Sub-harness | OpenCode 1.18.3 | Headless agentic loop w/ file + bash tools |
| Gateway | Vercel AI Gateway | One key, 0% markup, any model |
| Isolation | git worktree | Files sandboxed per worker |
| Guardrails | opencode.json | Deny-by-default bash allowlist |
| Instructions | AGENTS.md | Portable, agent-neutral |
The honest read
Pros
- Delegate any task to any model — any Vercel-gateway model is one registry line; any OpenAI-compatible endpoint, including local Ollama, is one config block.
- Own the harness — not locked to one vendor's UI.
- 0% gateway markup.
- Portable instructions — AGENTS.md is read by any harness.
- Every run logged with real cost.
Cons
- Claude subagents in-subscription still beat these workers on pure code ROI — this is an R&D + agnostic hedge, not a cheaper dev team.
- Workers run bash with the user's OS permissions — contained by a deny-by-default allowlist, not a full sandbox; a container/WSL jail is the next step.
- v1 handles Python / data / content tasks only — no Next.js / npm builds yet.
- Native-Windows setup had real friction.
Proof, not a pitch
7Worker runs
$1.25Total cost
1.18.3OpenCode
2026-07-17First pilot
4Frontier models
Case study: the first pilot run, graded by mutation testing →