What happened
DeepSeek’s GitHub release page lists v0.2.2-rc.2 as a pre-release, published on November 11. The release adds Git Worktree review organized by branch, commit, and file type. It also describes handling GitHub reviews, synchronizing review progress for files already viewed, and updating branches directly in DeepSeek Harness. The notes say the changes are intended to make reviewing changes easier within the app; they do not report measured review-time or defect-rate outcomes.
The other headline addition is support for named, multi-level SubAgents. AgentTeam can be enabled per session, with a shared member and task panel. The release also improves browsing and coordination for nested sub-agents: task descriptions receive priority, users can expand the hierarchy level by level, and failed reads can be retried. This gives teams a more explicit way to organize delegated work and inspect what each agent is doing, rather than treating sub-agent activity as an opaque sequence of calls.
The release notes include a desktop setting to switch between stable and development update channels. Other changes span code-review navigation and narrow-screen layouts, reduced lag when viewing large changes, and a longer wait allowance when creating worktrees in large repositories. The release also improves visibility into model context, including workspace instructions and skill catalogs, and makes agent terminals visible in the sidebar so a user can take over by typing.
DeepSeek Harness also changes some underlying workflows. The notes say archived sessions now clean up eligible Git worktrees by default while retaining branches and commits; users who need local directories preserved should disable cleanup. AgentTeam now uses unified named sub-agents, and custom plugins or SDK integrations may need migration. These details make the release relevant not only to end users but also to teams maintaining extensions around the harness.
What the archive adds
New Runtime covered DeepSeek Harness v0.2.1-rc.3 on October 24, then the v0.2.2 alpha with plugin updates and an experimental CLI sub-agent tool on October 28. This is a new release-candidate event in an ongoing product thread, with worktree review and named nested agents as the newly emphasized capabilities.
Trend signal
This release supports the view that agent products are becoming harnesses for coordinating and reviewing work, not just interfaces to a model. The AI-native organizations move toward review and orchestration pattern focuses on routing, verification, and ownership of agent output. The release provides product-level mechanisms for delegation and review, but does not show whether they improve accepted outcomes or reduce review time.
Why it matters
Delegation only helps a coding team if work can be divided without losing track of ownership or review state. Named, nested agents and a shared task panel make that structure more visible, while worktree review keeps changes separated and inspectable against branches, commits, and file types. Together, these features address two practical failure points in agent-assisted development: coordinating parallel work and checking what changed.
The release is also a reminder that the harness itself shapes the workflow. Teams adopting it should assess the review experience, nested-agent behavior, and extension migration costs—not just the model behind the agent. Because this is a release candidate, it is more suitable for evaluation than an untested production rollout. A useful trial would compare a representative multi-agent coding task with the team’s current process, recording review time, missed issues, and the effort required to integrate plugin or SDK changes.
What happens next
- DeepSeek Harness’s next stable release will retain named, nested sub-agent workflows and worktree review (3 months, 75%). Watch for: a stable release containing these features or release notes describing their removal.
Limits and uncertainty
The release page labels v0.2.2-rc.2 a pre-release, so the features and migration behavior may change before a stable version. The notes describe capabilities but do not provide independent evaluations of review quality, agent coordination, or performance on large repositories. They also do not establish whether the workflow fits every GitHub review process or how much extension maintenance teams will face beyond the migration cautions listed. The published release notes have not been independently tested here.
Skill to build now
Take one small issue that can be divided into independent changes and test it in a disposable Git worktree. Assign distinct, named sub-agents to bounded tasks, then review their changes by branch and commit. Record where you needed human clarification, whether the shared task view made ownership clear, and how well the worktree review handled already-viewed files. This gives your team a concrete basis for deciding whether to evaluate the release candidate further.
Reader takeaway
DeepSeek Harness’s release candidate brings nested agent delegation and Git worktree review into the same desktop workflow. For teams, the important question is not simply whether it can run more agents, but whether named ownership and inspectable changes make the resulting work easier to verify. Treat v0.2.2-rc.2 as an evaluation target, not a stable dependency.