DeepSeek Harness update adds file handling and agent-team experiments

THE ESSENTIALS

The September 10 update belongs to the Harness 0.1.5 release-candidate series.
The interface adds general file uploads and previews for documents, code, images and HTML.
Experimental agent teams require explicit installation and are not enabled by default.

A model update and a workspace update

DeepSeek's September 10 Harness release brings V4.1 Flash into its agent workspace alongside broader support for files and continuing conversations with subagents. The official repository labels the relevant 0.1.5 builds as release candidates, a more precise description than a completed stable release.

Harness is the surrounding application that gives a model access to a workspace and tools. Updating it changes how users supply materials, inspect generated files and direct ongoing work. Those changes are separate from the underlying model's training or benchmark performance.

Files become part of the conversation

The release notes make V4.1 Flash the default for new sessions unless a configuration explicitly selects another model. They add general file uploads, progress and cancellation controls, and a sidebar able to preview Markdown, code, HTML, PDFs and images. A model can hand over a file that users inspect in the interface or open with a local application.

This makes the output of an agent easier to examine as an artifact. A generated page or document no longer has to be represented only by a path in the chat. The workflow can keep the conversation visible while showing the actual deliverable alongside it.

Continuing work across agents

The changes also cover queued messages and steering for resumable subagents. Experimental Agent Teams packages remain opt-in. The notes describe support for changing a system prompt without invalidating the KV cache when the chosen model explicitly supports that behavior.

That qualification matters for long sessions: retaining cached computation is a runtime capability, not a universal property of every model connected to Harness. The same release sequence includes interaction fixes around pausing work and controlling busy sessions, reflecting the practical difficulty of keeping a long-running agent responsive to its user.