DeepSeek open-sources Harness, its plugin-based agent workbench

The same AI model can look brilliant in one setup and unreliable in another. On August 6 and 11, Composio placed DeepSeek V4-Flash inside eight agent harnesses — the execution systems wrapped around models — and asked each one to complete 30 multi-step tasks in Gmail, Google Calendar, GitHub and Slack. The best harness passed 20 tasks; the worst passed 14. Across 240 runs, only 129 succeeded, while just six tasks were completed by all eight systems.

DeepSeek’s black-whale team has surfaced Harness, a developer preview released under the MIT license on August 13. The project is designed as a workbench for putting models to work in real software environments, rather than as a finished agent sold as a single package. DeepSeek also released DeepSeek V4 Pro that evening.

The core design is that everything is a plugin. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling and the interface can all be replaced or recombined through the Cordis plugin system. During its internal preview, developers reportedly created hundreds of plugins within days, including one that adds cross-session long-term memory with background self-evolution. Harness defaults to supporting nearly 40 large models, including Kimi, OpenAI, Anthropic and Google systems.

The other major feature is a paper trail for autonomous work. Harness uses append-only session logs, recording what the model sees, system prompts, chain-of-thought, tool calls and results, sub-agent scheduling and every context injection. Compression does not erase the original history. In its Trajectory view, developers can restore, fork, retrieve and replay a run instead of trying to reconstruct why an agent failed from a final answer alone.

So what changes in practice? Developers can inspect the full path behind a success or failure, while models, tools, rules and interfaces can be swapped or recombined through plugins. Composio’s test put the stakes in concrete terms: Claude Code, Codex and DeepAgents each completed 16 tasks, but the cost was about $0.195, $0.081 and $0.045 per successful task. Harness is still marked v0.1, and the figures come from reported tests rather than an independent evaluation.