Extreme Harness Engineering for the 1B token/day Dark Factory — Ryan Lopopolo, OpenAI Frontier

Zero human code mandate

The team enforced a constraint of writing zero lines of code themselves, forcing the agent to become isomorphic to an engineer's capabilities and resulting in 1M+ lines of agent-written code.

10x speed inversion

Despite being 10x slower initially, the agent-driven approach became 10x faster than manual development, producing 1500+ PRs with just three people over five months.

Infinite parallel capacity

With 1 billion tokens processed daily, the system leverages massive GPU parallelism to work on unlimited codebase sections simultaneously, making human attention the only bottleneck.

Sub-minute build mandate

When GPT-5.3 introduced background shells making agents less patient, the team rebuilt the system from Make to Bazel to Turbo to NX to ensure builds complete in under 60 seconds.

Inverted environment architecture

Rather than pre-configuring environments, the agent serves as the entry point and spawns its own dependencies, using high-level tools like MI to boot observability stacks on demand.

Model-rev-driven refactoring

The codebase underwent five major build system changes in five months (GPT-5 through 5.4) as model capabilities evolved, requiring constant gardening of build invariants.

Post-merge human review

Humans moved from pre-merge review to post-merge oversight, accepting that synchronous human attention is the only fundamental scarcity while agent capacity is trivially parallelizable.

Autonomous agent merging

Agents autonomously merge code after review, with humans acting as circuit breakers only for critical issues rather than serving as manual gatekeepers for every change.

Systems thinking over debugging

The team focuses on building confidence in automation by constantly asking where agents make mistakes and encoding fixes into durable process documentation rather than fixing individual bugs.

Markdown-based steering

Used spec.md, agent.md, and skills.md files to guide behavior, with quality scores and tech debt tracked in markdown tables that agents review and update themselves.

Bidirectional review protocol

Code review agents and authoring agents can push back or defer feedback using priority frameworks (P0 vs P2), preventing non-convergent loops and scope creep from over-eager instruction following.

Durable process encoding

When fixing issues like missing timeouts, agents update reliability documentation to encode 'what good looks like' for future iterations, creating self-improving guardrails.