The center of gravity in Agent engineering is sliding from "model capability" to "how the model does not get stuck in real environments." DeepSeek packaged that idea into open source in early September: split the Agent into Model + Harness, where the model thinks and the Harness handles tool calls, session management, sandbox integration, and sub-agent scheduling.

Everything is a plugin: the Cordis kernel

The Harness code is open on GitHub (github.com/deepseek-ai/deepseek-harness). Its core is the Cordis kernel, a meta-framework that turns plugin mounting, unmounting, and dependency management into primitive operations. Every capability (models, tools, skills, sessions, sandboxes, storage, loops, scheduling, UI) is a Cordis plugin; you can swap or recompose anything in config, no source changes needed.

The academic foundation is arXiv 2608.25512 ("A Programming Paradigm for Spatiotemporal Composability", submitted August 26, 92 pages), authored by DeepSeek-AI and Peking University. The paper formalizes two orthogonal composition dimensions: temporal composability (a component's side effects can be fully reverted on removal) and spatial composability (inter-component dependencies can be declared and reactively managed). A unified "context type" mediates effects and coeffects, giving components safe interleaving without disturbing each other.

Four runtime modes, pick the Agent by use case

  • Standard: a full coding agent — file editing, shell, file and Web search, skills, planning, goals, sub-agents, workflows — all on. This is the default.
  • Code: tools are exposed as TypeScript via the Code Mode SDK; the model composes multi-step operations into a single program, reducing tool roundtrips.
  • Minimal: only bash and str_replace_editor remain — built for "benchmark models in a minimal environment."
  • Creator: built on Standard plus runtime inspection, plugin experiments, and preset authoring — for developers building custom Agent presets.

These four tiers are DeepSeek's answer to "how complex should an Agent be" — from a minimal shell for benchmarks, to a full production toolchain, to a meta-layer for builders.

Observability as the hidden sell

Every run enters an append-only session log: system prompts, reasoning, tool calls and results, sub-agent scheduling, context injections — every step indexed by source. The Trajectory view supports resume, fork, search, and replay over the same event stream. This turns the Agent's "think—act" trajectory into a replayable structured log, which makes model behavior debugging, offline evaluation, and training-data backfill noticeably easier.

So what?

DeepSeek is not the first to ship an open-source Agent harness (think Databricks Omnigent, the Anthropic SDK, LangChain-style frameworks), but it is one of the few that articulates Agent = Model + Harness most clearly, treats "plugin composition" as a first-class citizen, and ships observability on by default. Combined with the formal treatment of composability in the arXiv paper, teams whose Agents get stuck at the environment-interaction layer should look at Cordis's effect/coeffect design before reinventing the wheel.

References: DeepSeek Harness developer preview, arXiv:2608.25512, GitHub: deepseek-ai/deepseek-harness