"It's not a harness. It's CONSUME."
I'm a solo architect. I designed an agentic operating system in which the AI model is a replaceable engine, the ecosystem is raw material, and autonomy answers to a mechanical conscience. One person, working with Claude (Anthropic) — no team, no funding. Everything below is published, measured, and reproducible.
Most agent platforms ship a harness — an environment that optimizes agents inside it. My architecture inverts the relationship: harnesses, models, frameworks, and the open-source ecosystem are consumed — one-way, through a license-and-value intake gate, rebuilt as owned modules, run behind human gates (an autonomy dial where high-risk actions queue for a human even at maximum autonomy), with plain files as memory and sealed boundaries on the way out.
The proof is the record:
| Measured | |
|---|---|
| 10.533 billion tokens of engineering in 19 recorded build days (95.6% cache-read) | system's own meter |
| 165 capability modules · 247 logged build-and-verify iterations · 91 permanent failure immunities | live substrate walk, method published |
| 2026-06-11 — a provider announced a billing change; my 9-agent fleet was running on a different AI engine the same day, verified live 9/9 | the BYOM knob, under a real deadline |
| 2026-06-12 — consumed the 213K-star harness flagship: all 64 of its agents, license-verified, converted into my architecture's schema, every test green | the thesis, demonstrated |
agentic-os-audit — the professional engineering assessment
The architecture at pattern level, an implementation guide for teams, five reproducible simulations (stdlib Python, deterministic, one command), the full-data edition with every number measured live from the real system — and "Fable, Unsealed": the AI's own unsoftened opinion of the system it runs inside, weaknesses included, published exactly as written. Every claim labeled VERIFIED / PARTIAL / SIMULATED / MEASURED. No superlatives, by standing policy.
servari-open — the open reference shell (Apache-2.0)
A dependency-free, gate-first, bring-your-own-model agent OS: pure standard-library Python server, autonomy dial L0–L5 with the always-gate-high-risk floor, allow-listed action runner, file memory, append-only audits. A stranger's clean-room run passes 8/8 verification, 160/160 tests. Branch consume-ecc holds the consume demo — the harness flagship as raw material, opt-in.
The architecture, the CONSUME stance, and the autonomy-policy design are mine. The implementation, the assessment, and the simulations were written, tested, and mechanically verified by Claude (Anthropic) under my human-gated workflow — the same gate layer that blocked Claude from publishing this work until I said push. I trained a behavioral layer on top of a frontier model without touching a single weight; the model's own account of that is in the repo, signed.
- Outcomes shared, build-method withheld — stated openly, as a competitive choice.
- Honesty is the deliverable. Cost figures are API-equivalent cost-to-serve, not cash. Counts are inventories until validated. Unfinished tracks are labeled unfinished — the fine-tune regression stays in the record.
- Gates over trust. The actions that bite — deploy, send, spend, publish — always wait for a human. The gate is mechanical, not a prompt instruction.
📍 MEKreflect · mekreflect.com · honest engineering, on the record