ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
News2 min read·2026-10-21DisconfirmAI Editorial

Tenex Special Projects Lead CJ Hess: Enabling Agents to Read 'Why' Rather Than Just 'What Was Done'

At the AI Engineer New York conference, Tenex’s CJ Hess introduced MetaHarness: a process that commits plans, decision rationales, and acceptance criteria alongside code in Git, allowing agents to read 'why things were done' rather than just 'what was built.' This approach addresses a core barrier to enterprise AI agent adoption—context scattered across human brains and conversations instead of existing as retrievable, persistent records.

ImportanceMaterialEvidenceE3 inspectableWrite-upStandard

Original publication time 2026-10-21 12:52 GMT+8

At the AI Engineer New York conference, CJ Hess of Tenex described a dilemma engineering teams repeatedly face when deploying AI agents: agents can write code, but they don’t know why the team chose one architecture over another, why certain proposals were rejected, or what requirements clients changed. This information exists in conversations and engineers’ minds, not in the repository.

Tenex’s solution is called MetaHarness. CJ explained that the team initially tried standardizing a set of skills, which evolved into a structured documentation system within the repository accompanied by command-line tools. The core operation involves committing plan documents, acceptance criteria, and decision logs to Git alongside the code. When an agent returns to a task, it reads not only what was built but also why the team chose to build it that way.

"Code can be the memory of a codebase, but I think you still need documentation," CJ said in the discussion. "You need to know why that decision was made; you need to know the context of the conversation." He added a key insight: this information should "never be in someone’s head; it should always be in the repo." This gives agents and tools access to context that most engineering teams still keep only in conversations or human memory.

Sign in / Sign up
ReadingResearchRadarInvestment framework

Latent Space editors distilled a distinction from the episode: making agents write code faster and making the surrounding work legible are two different things. The latter includes plans, acceptance criteria, decisions, and evidence. Tenex’s own article on its AI-native SDLC also notes the cost: such a high degree of structure may be overkill for small teams working on a single project.

The conditions for applicability are noteworthy. Alex mentioned a principle during discussions on enterprise agent deployment: "You cannot automate a process that does not exist." If forty people roughly perform the same function but each does it slightly differently, building an agent to execute tasks in the background becomes more difficult. MetaHarness presupposes that the team already has some form of standardization—establishing the process first, then handing work to the agent, rather than the reverse.

Data from Tenex’s customers reveals a gradient in enterprise AI adoption: approximately 90% of surveyed enterprises have deployed general large language models to employees, about 60% have conducted training and capability building, and around 50% have deployed coding agents to engineering teams. However, only about 5% of enterprises have put agents into production for non-engineering work. Alex believes the biggest gap lies not in model capabilities but in permission management and process standardization—the exact types of problems MetaHarness aims to solve.

Sources

  • Source materialyoutube.com↗
Continue researching →