The model is commoditised. The harness is not.
HARNESS.md is designed to be parsed and run by an engine you own, once that engine (Drivetrain) is built. It does not reinvent existing standards; it composes with them. A HARNESS.md references AGENTS.md for project conventions and references LOGIC.md for each agent's reasoning contract.
You are building the runtime layer that the big labs only ship as closed monoliths.
Three composable layers. Clean separation of concerns.
The relationship is Dockerfile-to-app: the harness is designed to wire up and run the agents whose reasoning is specified in LOGIC.md. HARNESS.md is deliberately not folded into LOGIC.md, bolting orchestration, sandboxing, and runtime state onto a reasoning-contract format turns the flagship into a kitchen sink. And it's deliberately not a standalone brand, a separate identity fragments the ecosystem and walks into a crowded concept with no inherited equity.
Orchestrator, leads, workers. The fallback rule is the point.
A minimal HARNESS.md: one orchestrator, one lead, two workers, a small task DAG, one quality gate, and a references: to a LOGIC.md contract for each agent.
Orchestrator
Plans and delegates. Never builds. One human input in, internal fan-out only.
Leads
Plan and delegate to workers. Step in and do the work themselves when a worker fails, the fallback rule.
Workers
One agent, one prompt, one purpose. Strict domain scoping.
---
harness: leadscraper-enrich
version: 0.1.0
includes:
- AGENTS.md
teams:
- id: enrichment
lead: enrich-lead
workers: [scraper, normaliser]
agents:
orchestrator:
role: orchestrator
reasoning: ./contracts/orchestrator.logic.md
tools: [delegate, read_state, write_state]
enrich-lead:
role: lead
reasoning: ./contracts/enrich-lead.logic.md
tools: [delegate, fs.read, fs.write, db.query]
scraper:
role: worker
reasoning: ./contracts/scraper.logic.md
tools: [web.fetch, db.write]
models:
primary: anthropic/claude-sonnet
secondary: nvidia/meta-llama-3.3-70b-instruct
rotation: [primary, secondary]
---Illustrative sketch from the Phase 0 target, not final schema.
Reasoning and pushback, not just a conclusion.
Relationship to LOGIC.md: complementary layer, one stack
HARNESS.md is not folded into LOGIC.md and is not a rival standalone. It's the orchestration/runtime layer above it. The win condition: HARNESS.md drives demand for LOGIC.md, you cannot fill a harness without reasoning contracts, so every HARNESS.md adopter becomes a LOGIC.md adopter.
First deliverable: a walking skeleton, dogfooded
Not a spec document, not a full engine. A thin vertical slice touching orchestration, tool bindings, and control loop, run end-to-end on one real internal task with a captured trace. A spec with no executor is one more markdown convention in a crowded field. The differentiator is the runtime and the proof it works.
Stack: Python core, typed and enterprise-grade
Python 3.12+, Pydantic v2 for every contract, Pydantic AI as the per-agent leaf-execution primitive. The orchestration layer above it, control loop, shared state, failure recovery, eval and observability, is owned outright, not wrapped around LangGraph or CrewAI. That layer is the actual product.
Audience: both, open-source-first
Build private, dogfood on a real project, harden, then open-source as the HARNESS.md layer of the LOGIC.md stack, the same path LOGIC.md took, validated through Modular9 before it earned stars. Working tools that solve a real problem earn adoption; empty repos with grand READMEs don't.
Leadscraper: an existing pipeline, re-expressed as a harness.
Already a six-stage pipeline (Discover, Enrich, Analyse, Score, Assets, Outreach) that maps almost one-to-one onto orchestrator, leads, and workers. Already has human approval gates and a multi-provider LLM abstraction, exactly the model-rotation surface HARNESS.md needs.
Re-expressing the pipeline as a HARNESS.md-orchestrated team, fixing a known dedup bug, and clearing a backlog of 1,733 unenriched records is the demo. The proof and a real business win are the same artifact.
The walking skeleton, brief by brief.
| Brief | Scope | State |
|---|---|---|
| 001 | Drivetrain package scaffold + HARNESS.md Pydantic v2 schema + parser + drivetrain validate CLI |
Drafted, ready to execute |
| 002 | Control loop + fake model provider; orchestrator/lead/worker fan-out; induced worker failure → lead takeover → orchestrator reroute; emit tasks.json + progress.md | Not yet drafted |
| 003 | Real multi-provider model layer + Leadscraper enrichment dogfood; clear a batch of the unenriched backlog | Not yet drafted, operator-gated |
HARNESS collides with Harness.io, a funded DevOps/CI-CD company now shipping its own "Harness AI Agents." The file convention (HARNESS.md, generic and descriptive) is expected to stay; the runtime brand is Drivetrain to keep the product name clean and ownable, decided precisely to route around that collision.
Strategy locked. Engine not yet built.
The runtime layer of the LOGIC.md stack, alongside COVENANT.md and MARCHESE.md. This page gets rebuilt once there's something real to show.
Evidence: compiled from the harness-md repository on 2026-08-03. Every number on this page traces to a file path or command output in the source tree. Last updated: 2026-08-03.