In recent years, competition among large language models (LLMs) has centered on parameter counts, training data, inference capability and benchmark scores; but in the second half of 2026 a growing shift has emerged: the model itself is becoming one core component within an Agent system, while the other half that determines whether a model can truly complete complex tasks is the Agent Harness built around the model. DeepSeek recently open-sourced DeepSeek Harness, and OpenAI has been open-sourcing the Codex CLI and publishing its Codex Harness and App Server architecture. This signals that large-model competition is expanding from “whose model is stronger” to “who can build a stronger model runtime environment.”

It is important to clarify that the way Codex Harness was open-sourced differs somewhat from DeepSeek’s approach. DeepSeek published a standalone DeepSeek Harness project under the MIT license; OpenAI included the core Agent logic of Codex Harness in the open-source Codex CLI repository, which is licensed under Apache-2.0. OpenAI states that Codex Core handles both the Agent Loop runtime and single-thread persistence management, and exposes unified capabilities to CLI, IDE, web and other clients via the App Server.

The defining feature of DeepSeek Harness is “everything as a plugin.” Its underlying Cordis kernel is mainly responsible for plugin loading, unloading and dependency management, and models, tools, Skills, Sessions, Sandboxes, Storage, the Agent Loop, scheduling and even the UI are designed as composable, replaceable plugins. Developers can alter an Agent’s capability mix through configuration without modifying the Harness’ core code. DeepSeek also offers Standard, PTC, Minimal and Creator run modes; the PTC mode allows a model to generate a TypeScript program that combines multiple tool calls into a single program execution, reducing the frequent round-trips of the traditional “model–tool–result–model” pattern.

Another notable design is Trajectory, the complete Agent execution trace. DeepSeek Harness writes system prompts seen by the model, tool calls and results, context injections and sub-Agent scheduling into an append-only session log, supporting resume, branching, retrieval and replay. This turns an Agent from a one-off chat program into a stateful software system with history and debuggable execution traces.

By contrast, Codex’s approach emphasizes standardization across the Agent Loop, execution environment and clients. OpenAI positions the Codex Harness as the common runtime behind all Codex products, including web, CLI, IDE and desktop apps. Codex Core is responsible for the Agent Loop, thread lifecycle, persistence, tool execution and extensibility, while the App Server exposes shared Agent capabilities to different clients via a bidirectional JSON-RPC interface.

Also noteworthy is OpenAI’s recent concept of “Harness Engineering.” Internal experiments indicate that in projects entirely generated and maintained by Codex, human engineers’ primary work shifts from “writing code” to designing environments that Agents can understand and verify, including clear code structure, AGENTS.md, structured documentation, automated tests, CI, logging, metrics, browser control and executable feedback loops. OpenAI even reports that one internal product reached roughly one million lines of code in about five months and that team engineering efficiency improved significantly. This disclosure reflects OpenAI’s own engineering practice and is not an independent third-party benchmark, so it should be viewed as evidence of a trend rather than a universal performance claim.

The two architectures actually represent two complementary paths: DeepSeek stresses the Harness itself as modular, composable and e