DeepSeek Harness has published the first release candidate for version 0.2.0, adding practical fixes for stalled conversations and clearer recovery when a tool’s result is uncertain. The update also lets conversations using DeepSeek account models search the web without an additional API key, alongside a spread of desktop and workflow improvements.
DeepSeek Watch analysis
What happened
The release candidate, published on 28 September, gathers changes since v0.1.7-rc.2. DeepSeek Harness says an operation with an unknown outcome now prompts the agent to check for side effects before retrying, rather than blindly repeating the action. It also fixes conversations becoming stuck after a tool-scheduling failure.
Other changes include automatic re-upload of expired images, improved Office and PDF text selection, and fixes for desktop window layouts, macOS microphone permissions, Safari response recovery and some Linux installation failures. The release notes also describe a Windows sandbox skill for diagnosing supported permission problems and making backed-up, recoverable repairs within an authorised directory. DeepSeek Harness 0.2.0-rc.1 release notes
Why it matters
The most consequential change is about what an agent does after something goes wrong. A tool call that may already have changed a file, sent a message or triggered another action is a poor place for automatic do-overs. Asking the agent to verify the side effects first is a useful guardrail, though the notes do not claim it prevents every mistaken retry.
For users, the release also chips away at everyday friction: fewer broken sessions, clearer recovery and web search without another API key for DeepSeek-account conversations. This is a release candidate, so it is a preview of the 0.2.0 series rather than a final stable release.
Our read
There is good, unglamorous judgement in making an agent check what happened before it tries again. That matters more than a freshly polished menu when software can take actions on a user’s behalf. If you use the release candidate, treat it as a preview and pay particular attention to the boundaries around those recoverable Windows permission repairs.
What to watch
- Whether the final 0.2.0 release keeps these changes and adds further fixes.
- How the side-effect check behaves when a tool’s outcome is genuinely ambiguous.
- Which account and configuration requirements apply to the new web-search capability.
- Whether users find the Windows diagnosis and recovery flow clear and predictable.
Discussion spark: When an AI agent is unsure whether an action succeeded, should it stop and ask every time, or is checking side effects itself a safe enough default?
Sources and evidence
- v0.2.0-rc.1 (28 September 2026, 12:36 UTC)
not affiliated with or endorsed by DeepSeek