llms.blog
AI

DeepSeek Open-Sources DeepSeek Harness: Modular Agent Runtime Built on Cordis Plugin Framework

DeepSeek AI has released DeepSeek Harness (dsh), an open-source agent harness built on a modular "everything is a plugin" design. Published under the MIT license on GitHub, the project provides a runtime framework where all core agent functions—including model adapters, session storage, tool registries, sandboxes, and the execution loop itself—are implemented as interchangeable, hot-swappable plugins. Rather than developing a monolithic agent runner, DeepSeek structured the system around Cordis

axc
axc
2 min
DeepSeek Open-Sources DeepSeek Harness: Modular Agent Runtime Built on Cordis Plugin Framework

DeepSeek AI has released DeepSeek Harness (dsh), an open-source agent harness built on a modular "everything is a plugin" design. Published under the MIT license on GitHub, the project provides a runtime framework where all core agent functions—including model adapters, session storage, tool registries, sandboxes, and the execution loop itself—are implemented as interchangeable, hot-swappable plugins.

Rather than developing a monolithic agent runner, DeepSeek structured the system around Cordis, an open-source composition meta-framework originally authored by developer shigma and refined over four years within the Koishi chatbot ecosystem.

DeepSeek Harness Runtime Architecture and Cordis Plugin Lifecycle

The Cordis Foundation and Reversible Effects

DeepSeek Harness leverages Cordis v4 to solve key lifecycle and state management challenges common in dynamic agent environments:

  • Reversible Side Effects (ctx.effect): Every plugin modification registers an explicit inverse cleanup function. When a plugin or tool is unmounted or reloaded, Cordis automatically executes its teardown in reverse registration order, preventing resource leaks and dangling state.
  • Spatial and Temporal Composability: The framework formalizes dependency management, ensuring plugins only initialize after prerequisite services are active and unload before upstream dependencies terminate.
  • Path Independence: Application state is determined strictly by the set of currently enabled plugins rather than the sequence in which they were attached or modified.

Runtime Modes and Multi-Agent Orchestration

DeepSeek Harness ships with a local Web UI (defaulting to port 3080), headless execution modes, and a Python SDK for programmatic integration. The environment supports four distinct session presets:

  1. Standard Mode: A full-featured agent environment including file editing, workspace terminal execution, web search, skills, planning tools, and subagent delegation.
  2. Code Mode: Executes multi-step workflows by allowing the model to author single TypeScript programs against a unified SDK, collapsing multiple sequential tool invocations into a single execution turn.
  3. Minimal Mode: A stripped-down configuration containing only a persistent bash shell and a file editor, replicating the baseline environment DeepSeek uses for standardized evaluations.
  4. Creator Mode: An inspection and authoring preset allowing developers to experiment with plugin configurations and author custom session templates.

The harness natively supports parallel subagent orchestration, providing live progress indicators for concurrent workers and a detailed Trajectory timeline that profiles turn latencies, token consumption, and granular tool execution metrics. Provider connectivity is managed via the pi-ai library, supporting official DeepSeek endpoints, OpenCode Go, and generic OpenAI-compatible APIs with credential segregation in local configuration files.

DeepSeek AI currently designates the repository as a developer preview with active community contributions and plugin development underway.

Sources

AIOpen SourceLLMAgentsnews
ShareX (Twitter)LinkedIn

Written by

axc

More to read

  • Hallucination Detection and Faithfulness Verification in Production RAG: Architecture, NLI Claim Decomposition, and Runtime Guardrail Economics

    Retrieval-Augmented Generation (RAG) is commonly deployed under the assumption that grounding generation in retrieved passages eliminates factual inaccuracies. In practice, grounding provides an evidence boundary but does not guarantee factual fidelity. Production language models regularly synthesize claims absent from the retrieved context (extrinsic hallucinations) or directly assert statements conflicting with retrieved premises (intrinsic contradictions). As enterprise RAG pipelines scale i

    1 min

  • Sparse Attention and BigBird: How Window, Global, and Random Graphs Preserve Turing Completeness in Linear Time

    Standard self-attention in transformer architectures scales quadratically with sequence length. Computing full pairwise interactions between n tokens requires evaluating an n x n attention matrix, yielding O(n^2) computational complexity and memory consumption. While hardware accelerators and IO-aware tiling algorithms like FlashAttention optimize memory traffic, the quadratic compute and KV footprint remains a barrier for processing long contexts, document-level summarization, and genomic seque

    1 min

  • Oxford Study Details Chinese Gray-Market Proxies Reselling Claude Tokens at 90% Discounts

    An investigation by the Oxford China Policy Lab reveals that Chinese developers routinely access Anthropic's frontier Claude models at discounts between 70% and 90% below list price, bypassing geographical blocks, payment filters, and biometric identity verification through a decentralized network of API proxies known locally as "transfer stations" (中转站). The analysis, authored by Oxford researcher Zilan Qian and published via ChinaTalk, outlines the modular supply chain and economic mechanics

    1 min