DeepSeek Harness vs. Codex
DeepSeek Harness asks a useful question: what changes when the agent harness—not only the model—is programmable all the way down?
The short version
DeepSeek Harness (DSH) is an open-source agent harness in developer preview. Its headline is literal: every major runtime capability is a plugin. That includes the parts most agent products hide behind a fixed interface—sessions, sandboxes, execution loops, scheduling, storage, and the UI.
Codex and DSH are therefore not clean substitutes. Codex is an integrated environment for getting difficult work done with an agent. DSH is especially interesting to people building the environment itself: teams who want to exchange components, inspect trajectories, or test a radically different agent loop without forking an entire product.
Everything is a plugin is a systems claim, not a packaging detail.
DSH uses the Cordis kernel to mount plugins and manage their dependencies. The consequence is architectural: a team can change a capability by composing configuration, rather than modifying the harness source. That pushes the boundary of experimentation below prompts and above infrastructure.
Two useful views of an agent runtime.
The unit of customization
DeepSeek Harness
A capability is a plugin. The model, tools, sessions, sandbox, storage, loop, scheduler, and UI can be selected, swapped, or extended in configuration.
Codex
Codex is the polished agent experience: a capable model working through a deliberate set of tools, skills, permissions, workspace context, and collaboration controls.
Harness treats the agent runtime as the product. Codex makes the runtime legible enough that a developer can focus on the work.
The unit of evidence
DeepSeek Harness
Its append-only session log records the system prompt, reasoning, tool calls and results, subagent scheduling, and context injection. Resume, fork, search, and replay share that event stream.
Codex
Codex preserves the working conversation, files, tool outputs, and task state so the human and agent can collaborate across a real workspace.
Both value durable work. Harness makes the trajectory itself a first-class, reconstructable runtime primitive.
The unit of experimentation
DeepSeek Harness
Standard, Code, Minimal, and Creator modes expose distinct operating envelopes—from a full coding agent to a two-tool benchmark environment and an in-memory plugin laboratory.
Codex
Codex concentrates on productive software work, while skills and workspace instructions provide reusable, scoped specialization.
Harness makes agent-environment design unusually easy to vary. That is powerful for researchers and platform teams who need controlled comparisons.
Trajectory as infrastructure
Recording every context injection alongside actions and results gives debugging a stronger object than a chat transcript. If a run is reproducible from one log, production failures can become test cases rather than anecdotes.
Minimal mode as a control group
A persistent shell and a file editor are intentionally sparse. That makes it easier to separate model capability from the scaffolding around it when evaluating agent behavior.
Code mode as agent-side orchestration
Instead of only making sequential tool calls, the model can write a TypeScript program to combine multi-step operations. This reframes code generation as a control plane for work, not just the work product.
Creator mode as a meta-tool
Runtime inspection, in-memory plugin experiments, and preset authoring make the harness capable of helping construct its own variants. That is a promising path toward disciplined harness engineering.
The lesson is not to make every product configurable.
Most teams should prefer a narrow, dependable agent environment over a universe of switches. But DSH makes an important contribution: treat the harness as an explicit engineering surface, preserve the evidence of every run, and make experimental changes cheap enough to evaluate honestly.