Breaking

InstructGPT Co-Author Diogo Almeida Publishes PDF on Building a 'Jev Harness' for Coding Agents

November 9, 2026 at 13:16 EST

AI Agents
Software Dev & Coding

Diogo Almeida, a co-author of the InstructGPT paper, has published a PDF describing how to build a 'Jev harness' for AI coding agents, and he argues that the surrounding system, rather than the underlying model, is what sets products apart.

AI coding agents · Harness design

The harness, not the model, may be the real moat

InstructGPT co-author Diogo Almeida has published a PDF on building a "Jev harness" for AI coding agents, arguing that the surrounding system sets products apart more than the underlying model does.

How the approach is meant to be used

STEP 1

Harness PDF

Describes the scaffolding around a model

→

STEP 2

Existing agent

Claude Code or Codex reads the document

→

STEP 3

Rebuilt setup

The agent configures its own working environment

Core claim

"The model is not the moat. The harness is."

A harness is the scaffolding that shapes how a model behaves in practice. As frontier models become more widely available, the argument goes, competitive advantage shifts to how teams wrap and direct them.

Instructions Tools Workflows Feedback loops

What supports the approach

Built by a notable figure in the field
Designed for direct use with existing coding agents
Fits a broader shift toward tooling, context and orchestration

What remains unverified

No independent benchmarks or performance results
No public detail on how the harness works
No third-party evaluation found so far

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.