Atlas Harness

Atlas Harness is an operating desk for AI work. The portfolio dates it July 2026 to the present, in TypeScript and Python, formerly codeharness, with 406 spine commits. It is a policy and recovery layer for coding tools. The desk defines who may act, when work is complete, which lane owns the work, and what evidence remains.

The bird's-eye map, covering 18 June to 13 August 2026, puts the harness row in public view. On 25 July 2026 the map marks codeharness and four primitives, with the line "model proposes, harness disposes." On 26-27 July it marks the rename to Atlas as the sole home, with the label "codeharness name dies." The same row marks desktop and remote work, Gate A closed on 1 August, a board on port 4319, and the Boston atlas CLI as a front door on 10 August, beside a separate July Tauri desk. A GitHub leftover on the map is labeled with history from 25 July to 1 August.

Jul 2026 onward, in the portfolio story, is when the operating desk was built. Atlas Harness turns the earlier prose contract into typed policy, clear work lanes, completion checks, source records, recovery, and evaluation. Atlas Symphony is the named process that runs that desk.

What the desk enforces

The coordination pattern is steward, heartbeat, workers, review team, and historian. The topology is enforced.

Tool interfaces treat Claude, Codex, and Grok as engines on existing subscriptions.

Human oversight uses typed completion gates so final approval stays with the operator.

The project tree table lists Atlas Harness as the operating desk for policy, lanes, reviews, and evaluation: 406 commits and 118k lines.

The evidence audit found 64 Atlas Harness pull-request records, including 62 merges. The retrieved work covers durable jobs, provider adapters, an operator board, evidence controls, an Obsidian knowledge base, an encrypted Context Vault, recovery source, governance, and personal-model lifecycle controls.

Beside the chat mind

On the Atlas page, work orchestration is a separate plane from conversation. For coding work, Atlas sits above live-probed Codex, Claude, and Grok runtimes. Tasks enter a protected mission lane, continue through bounded handoffs or failover, and remain attached to an isolated worktree, a persistent event history, and a computed done gate.

Queen is the separate heavy-lift lane. Hive 4.0 extends that work through temporary fleets, fenced leases, freeze controls, and evidence rollups. J-space runs declared research scenarios and records the resulting responses.

The conversational mind, model servers, work harness, and research environment exchange context and evidence, while each keeps its own authority, persistence rules, and failure boundary.

Computed completion keeps source, checks, deployment, and live behavior as separate evidence gates. Human control keeps spending, residency, consent, shutdown, and real-world authority bounded.

The current source implements the linked turn pipeline, separate ledgers, model routing, session boundaries, sealed state engines, and work-harness controls described on the Atlas page.

Orchestrating GPU compute

I managed a system for two billed GPU seats from an ARM64 Pi.

GPU pods incur per-minute billing. The Pi control plane initiates billing only upon verified intent.

I implemented a hub-and-spoke architecture. The Pi CLI treats GPU pod status, health, and metadata as read-only probes, explicitly decoupled from the lifecycle `up` command.

This separation ensures deterministic billing control. The system supports reliable pod management on resource-constrained hardware with industrial stability.

The evidence finding states the same pattern: Atlas manages two billed GPU seats on ARM64 hardware through a hub-and-spoke control plane. The system treats all status, health, and metadata requests as side-effect-free probes, ensuring billing is only initiated by verified, intentional `up` commands.

In the working set, TypeScript and Python run across Atlas Harness, MobPay, MobHub, BeltBrain, research systems, and this site.