DeepSeek Harness v0.1.6-alpha.1 adds experimental Browser Use and Computer Use, giving the agent workbench new ways to interact with browsers and local machines. The release also adds terminals, MCP resource discovery, headless JSONL runs and remote-workspace support over SSH, making it a substantial step towards agents that can do more than chat politely about the work.
DeepSeek Watch analysis
What happened
The DeepSeek Harness release notes list Browser Use through Playwright MCP, Chrome DevTools MCP and Stagehand, plus Computer Use through Cua Driver MCP or a native driver that can operate a local computer and capture screenshots. Both features are labelled experimental.
Developers also get Web-sidebar terminals with multiple tabs, archived-session browsing and restoration, MCP resource discovery and URI templates, and headless execution that reads tasks from standard input, resumes sessions with –session-id and emits newline-delimited JSON with –json. DeepSeek says the release can run locally while using a remote workspace over SSH.
Our top picks
- Browser Use
Experimental Playwright MCP, Chrome DevTools MCP and Stagehand backends give agents a route into browser workflows. - Computer Use
Cua Driver MCP or a native driver can operate a local computer and obtain screenshots. - MCP SDK v2
Protocol negotiation, paginated tool lists and servers without tools make integrations less brittle. - Headless runs
Standard-input tasks, session resumption and JSONL events make automation easier to wire into existing systems. - Remote workspaces
DSH can run locally while working against a remote workspace over SSH. - Auto review
The release adds an experimental Auto review mode for another pass over agent output.
Why it matters
The useful shift is from an agent that operates mainly inside its own interface to one that can reach into browsers, desktops, terminals and remote workspaces. That makes the project more interesting for real engineering workflows, but also raises the stakes around permissions, failure recovery and what an agent is allowed to touch.
This is an alpha release, and the notes do not provide performance measurements or a safety evaluation for the new interaction modes. Treat the new controls as a test bench, not a trusted digital colleague with house keys.
Our read
DeepSeek Harness is becoming a proper agent workbench, with the connective tissue now as important as the model call. The release is worth exploring for developers who want browser or computer interaction, provided experimental means experimental and the permissions stay on a short leash.
What to watch
- Whether Browser Use and Computer Use become reliable beyond demonstrations.
- What permission and sandbox controls apply to local computer actions.
- Whether MCP SDK v2 breaks existing integrations during migration.
- How remote-workspace and headless modes behave in unattended runs.
Discussion spark: Which matters more for useful agents right now: reliable browser and computer control, or better visibility and recovery around the tools they already use?
Sources and evidence
- v0.1.6-alpha.1 (15 September 2026, 04:59 UTC)
not affiliated with or endorsed by DeepSeek