lopopolo/harness-engineering

What it solves

It addresses the gap between general model weights and an organization's private, evolving process data. While LLMs are capable, they lack the specific operational state, local ontology, quality bars, and internal procedures (the "process-data iceberg") that a worker needs to perform a specific organizational job reliably.

How it works

Instead of focusing on the model itself, this approach treats the model and coding agent as a black box and focuses on "harness engineering"—shaping the environment around the agent. This is achieved by:

• Curating Context and Tools: Providing the agent with the necessary external levers to recover intent and operate systems.
• Encoding Nonfunctional Requirements: Turning quality attributes (reliability, security, performance) and organizational decisions into retrievable context, examples, and executable constraints within the code repository.
• Cumulative Judgment: Creating a feedback loop where lessons from failures, corrections, and user responses are fed back into the environment to improve future agent trajectories.
• Last-Mile Deployment: Ensuring the organization's context, capabilities, and authority are supplied to the agent at the point of execution.

Who it’s for

Organizations and developers who use coding agents to maintain systems and want to ensure those agents adhere to specific organizational standards, quality bars, and internal operational knowledge.

Highlights

• Environment-Centric: Focuses on improving agent output by shaping the environment rather than the model.
• Nonfunctional Requirements as Code: Aims to get the whole universe of organizational constraints into the code.
• Cumulative Coherence: Makes organizational judgment cumulative so that agent-maintained artifacts remain coherent over time.
• Agent-Ready Documentation: Includes an AGENTS.md file to route agents to relevant arguments, cases, and proof.