AIGC · Agent

DeepSeek Harness Ships a Plugin-First Agent Runtime Where Every Capability Is Swappable

By 怕浪猫 · Aug 13, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Agent frameworks are proliferating, but most hard-wire their component choices. A plugin-first architecture that treats every capability as a swappable module lets teams mix their own models, tools, and sandboxes without forking the framework—useful when compliance, cost, or performance demands a different stack per deployment.

Summary

DeepSeek Harness v0.1 is an open-source agent runtime built on the Cordis plugin system. Its core design principle is "everything is a plugin": models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI are all independent plugins that can be swapped or extended without touching framework source code. Four preset modes—Standard, Programmatic Tool Calling, Minimal, and Creation—ship with different plugin bundles for different workloads, from full tool suites to bare-bones benchmarking. Every agent run produces an append-only event log that captures system prompts, chain-of-thought, tool calls, sub-agent dispatch, and context injections, all replayable and forkable from a single trajectory view. The project is in early developer preview and invites community plugin contributions.

Takeaways

— Every agent capability—model, tool, skill, session, sandbox, storage, loop, scheduler, UI—is a Cordis plugin that can be replaced or extended without modifying the framework.

— Four built-in modes bundle different plugin sets: Standard (full tools), PTC (model-generated code orchestrates tool calls), Minimal (shell + file editor for benchmarking), and Creation (inspect and compose new modes at runtime).

— An append-only session log records everything the model sees, including system prompts, chain-of-thought, tool calls, sub-agent scheduling, and context injections.

— The Trajectory view lets developers inspect, recover, fork, retrieve, and replay runs from a shared event stream.

— Installation requires Node.js; a quick-start Web UI launches with `npx @deepseek-ai/dsh web`.

Conclusions

Making the UI itself a plugin signals that the framework expects to be embedded or replaced in custom tooling, not just used as a standalone app.

PTC mode shifts tool orchestration from the model's function-calling interface to generated code, which can reduce token spend and improve determinism for multi-step tool sequences.

The Creation mode, which allows runtime inspection and in-memory plugin experimentation, treats the framework as a live programmable environment rather than a static configuration file.

Concepts & terms

Cordis plugin system

A meta-framework that manages plugin loading, unloading, and dependency resolution. Plugins communicate via services and events and can be composed at the configuration layer without code changes.

PTC (Programmatic Tool Calling)

A mode where the model generates code to orchestrate multiple tool calls in a single step, rather than making sequential function-calling requests. This can reduce round-trips and token usage.

Trajectory view

A UI that replays an agent's full execution trace from an append-only event log, showing system prompts, reasoning steps, tool calls, and context injections grouped by source.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗