DeepSeek Harness v0.1.0-rc.8 Brings Multimodal Inputs to Agent Workflows

Introduction

Only a week after entering public beta, DeepSeek Harness has released v0.1.0-rc.8. The release includes 14 updates, with multimodal input standing out as the most important change. Harness can now accept native image requests as well as mixed text-and-image input. This should not be read as a claim that an underlying text-only model has independently gained vision. More precisely, the harness now provides an input and orchestration path for images, allowing compatible models or backends to process them.

Key changes

• Native image and mixed-input requests: A request may include an image, text, or both, giving users more ways to describe a task.
• Images in workflow commands: Commands such as /goal and /plan can directly receive images. Screenshots and other visual references can therefore become part of a task goal or planning context.
• Further sub-agent integration: Claude Code and Codex are connected more deeply to the sub-agent system. The release summary does not specify scheduling rules or limits, so it would be premature to describe this as a fully defined multi-agent platform.
• Persistent PowerShell on Windows: A persistent PowerShell session reduces the need to repeatedly recreate a terminal environment.
• Minimalist mode enabled by default: The default experience is intended to be more streamlined, although the exact interface differences should be verified in the actual release.

Why it matters

Multimodal support has often been associated with a single chat turn: upload an image and ask for an answer. Harness moves that capability closer to an agent workflow. A screenshot can serve as part of a goal, visual information can influence planning, and an image does not have to remain an isolated attachment. That model is more relevant to coding, interface inspection, and terminal-oriented tasks.

The boundaries are important, however. Harness supporting image requests solves an integration and orchestration problem; it does not guarantee identical visual understanding across all backends. Actual image comprehension, supported formats, and output quality still depend on the connected model or service. The safest interpretation is therefore that rc.8 is an infrastructure update for multimodal agent workflows, not a new foundation-model release.

The rapid addition of multimodal input, sub-agents, and Windows session support shows that the project is improving input, execution, and developer-environment layers in parallel. Future releases will be worth watching for stable image context across planning and execution, clearer sub-agent responsibilities, and consistent behavior across model backends.

Source: OSChina