OpenAI's Codex harness reveals agent governance patterns teams should steal

OpenAI engineer Dominik Kundel presented the internals of the Codex agent harness at AI Engineer World's Fair 2026. The talk is not a product pitch but a technical blueprint. The harness is open source (MIT/Apache 2, written in Rust) and exposes two key protocols: the app server protocol for UI-to-harness communication and the responses API

for harness-to-inference. Both are designed as open ecosystems with third-party integrations like Ollama, LM Studio, and NVIDIA. The real signal is that OpenAI is trying to create a standard layer that other agent builders can adopt rather than building proprietary, closed infrastructure. The most operationally relevant design detail is context construction. The harness caps available

skills at 2% of the max context window and uses deferred tools via 'tool search' to avoid token bloat. This is a direct response to the practical problem that MCP plugins and toolsets grow unpredictably, increasing latency and cost. Teams building agents should study how OpenAI trades off context size for cashability and flexibility, as

this is the primary cost driver in production. The auto-review subagent for sandbox approval is the most consequential feature for enterprise deployments. Rather than granting full access or suffering approval fatigue, the harness spawns a read-only subagent to evaluate tool calls against user intent and risk taxonomies. This is a governance mechanism that addresses the