Bill Chen and Brian Fioca of OpenAI's Applied AI team present a methodology for building durable "Coding Agents" capable of surviving rapid model evolution cycles. They break down the anatomy of an agent into three parts: the User Interface, the Model (the raw intelligence), and the "Harness" (the integration infrastructure).
The core of their talk focuses on the underestimated complexity of the Harness. Building a robust harness is difficult because it must handle adaptation to each model's specific "habits" (prompt engineering), latency management, context compaction, tool integration (MCP), and above all, security (sandboxing). They point out that models have "personalities" or training biases (such as the tendency to want to read everything before acting) that the harness must manage in order to be effective.
To address this problem, OpenAI positions Codex not merely as a model but as an integrated "agent + harness" solution. Codex (the agent) natively handles the complexity of interacting with the system: running terminal commands, editing files, managing memory, and creating tools on the fly. This makes it a "Computer Use Agent" for the terminal, capable of carrying out tasks ranging from pure coding to system administration or data analysis (CSV files, folder organization).
They identify emerging patterns for agent developers: 1. Harness as an abstraction layer: Using a robust harness (such as the SDK Codex) allows teams to focus on product differentiation rather than on maintaining low-level infrastructure (prompting, tool calling). 2. Agents within Agents: Integrating Codex via SDK as a "sub-agent" capable of executing complex technical tasks within a larger application. 3. Dynamic tool creation: The ability of agents to write their own connectors (MCP) to interface with APIs or systems without prior integration.
In conclusion, they encourage developers not to "reinvent the wheel" of the harness each time, but instead to rely on existing infrastructure to build richer, more stable user experiences.