Harness Engineering
“Most people do not know that they can just point their agents at my writing, tweets, podcasts, and talks and improve the output of their agents by 100x.”
Harness engineering, the practice of improving agent output by shaping the environment around it, holds a chosen model and coding agent constant as a black box. It improves the two external levers—context and tools—and curates the environment around them. The worker should be able to recover intent, operate the real system, respect authority, prove the outcome, and leave the next run better equipped.
A central purpose of that environment is to carry an organization's nonfunctional requirements: the quality attributes and constraints governing reliability, security, compatibility, maintainability, performance, operability, risk posture, and polish. The harness also carries local decisions about how to prioritize, trade off, and satisfy those requirements. Ryan adopted a systems-level framing from 2026’s [un]prompted conference that describes this as getting the whole universe of nonfunctional requirements into code. Make the Repository Teach the Agent develops how the requirements and decisions become retrievable context, examples, tools, and executable constraints.
Because work is an iterative game, a harness can make organizational judgment cumulative. Lessons from accepted work, corrections, failures, and user responses become context, boundaries, tools, examples, and checks that shape later trajectories. Over time, that feedback loop can make coherence cumulative across agent-maintained artifacts.
Code is how an agent uses a computer. That internal action language can produce reliable domain outcomes for people who never review the implementation when last-mile deployment supplies the organization’s context, capabilities, authority, and proof.
General model weights contain only the visible tip of an organization’s process-data iceberg. Below the waterline sit the current operational state, local ontology, quality bar, procedures, exception history, and authority relationships that an agent needs to do a particular job. Organizations cannot presume that this private, changing process data will be present in general model weights, nor that agents will reliably intuit which process data matters to the organization. Harness engineering is the last-mile work of making it available to a capable worker as context and tools.
Point a coding agent at this repository alongside the system it should improve.
AGENTS.md routes the task to the relevant arguments, cases, and proof. For
direct reading, start with the thesis index. For an application, choose from
the playbooks.
Sources and related work
- “Harness engineering: leveraging Codex in an agent-first world” (fetch helper for agents blocked by the canonical page)
- Source library
- Influences and alternate framings
Repository-authored material is licensed under CC BY 4.0. See COPYING.md
for attribution and rights in source material.
