Unlock the Power of Multimodal AI with DeepSeek Harness
Just one week after launching its public beta on August 13, DeepSeek has shipped version v0.1.0-rc.8 of DeepSeek Harness, the open-source AI agent execution runtime that has drawn attention for its modular, plugin-based architecture. The update delivers 14 changes spanning multimodal input, subagent collaboration, terminal improvements, and developer tooling.
Multimodal Vision for Every Model
The headline feature is native multimodal support. Model adapters can now enable image requests through configuration, sending images directly to vision-capable models. The /goal and /plan commands accept mixed text-image input, and the input menu adds file and session references so developers can pull local screenshots into agent workflows.
More notable is how the framework handles models that lack built-in vision. When a direct image read fails, Harness falls back to OCR text recognition, color-ratio statistics, and partial pixel-row scanning, decomposing an image into structured information that a text-only model can reason over. Developer Yinsen found that screenshots of presentations, flowcharts, and UI layouts with clear structure recover useful information through this approach, though real photographs yield less. The community has already produced several vision plugins — including dsh-vision, dsh-vision-toolkit, and modlens — that now work alongside the native multimodal pipeline.
Subagent and Terminal Upgrades
The subagent system expands with support for installing Claude Code and Codex as Profile Bundles on demand, with Codex supporting non-interactive permission mode and multiple named instances. Windows users gain persistent PowerShell sessions in the PTY terminal, enabled by default in minimal mode. Bug fixes address oversized history images that caused failed requests, canceled streaming that lost reply prefixes, and custom OpenAI-compatible gateways failing on request-format differences.
A Modular Bet on Agent Infrastructure
DeepSeek Harness is built on the Cordis meta-framework and uses a micro-kernel architecture in which runtime components — model adapters, tool registries, sandboxing environments, session state handlers, and event dispatchers — operate as isolated, interchangeable plugins. The platform includes an append-only event log that records every user message, tool invocation, and reasoning state into a unified execution trajectory for debugging and benchmarking. Version 0.1 ships with four baseline configurations: Standard, Code, Minimal, and Creator modes.
As InfoQ noted, the release reflects a broader industry shift toward unbundled infrastructure for AI agents, though how widely the runtime is adopted will depend on the stability of its plugin ecosystem and long-term API maintenance as the framework moves beyond its preview phase.
