A look at OpenAI's open-source agent harness that now powers Codex and ChatGPT Work, as the company works to optimize the harness to cut runaway token usage
Context & Ripple Effects
OpenAI’s agent infrastructure has moved from a coding-specific product into shared workplace tooling: Codex began as a software-engineering agent, then Codex-powered shared workspace agents extended that model inside ChatGPT.
The harness now sits beneath both Codex and ChatGPT Work, making its efficiency consequential across a user base that OpenAI recently said had reached 10 million users of the two products.
First-order effects
- Optimizing the shared harness directly targets runaway token consumption in Codex and ChatGPT Work, reducing the cost and operational overhead of agent runs if the changes work as intended.
- A common harness means OpenAI can apply runtime improvements across its coding and workplace-agent products rather than maintaining separate execution layers.
Second-order effects
- Lower token use can make longer or more iterative agent tasks more economically viable for OpenAI and its customers, particularly where repeated tool use drives consumption.
- The move raises the bar for competing agent platforms: useful agent features increasingly need runtime efficiency alongside capable models and polished interfaces.
Third-order effects
- Agent competition is likely to shift toward agentic unit economics: the ability to control inference and tool-use costs may determine which products can sustain broad deployment.
- If shared harnesses become the standard control point, product differentiation may increasingly sit above a common governed runtime—through workflows, permissions, and integrations rather than a separate agent loop for every surface.
The trend: This is one data point in the maturation of AI agents from model demonstrations into shared work surfaces whose economics depend on disciplined runtime control.