dsh-agent-teams: coordinated subagents inside DeepSeek Harness

When AgentTeams is a useful fit

AgentTeams is designed for work that benefits from explicit roles and dependencies: parallel research, implementation plus review, multi-area code audits, and staged verification. The current session remains the captain and consolidates the result; members are continuable DSH subagents rather than an external workflow service.

The plugin provides a planning step before execution, durable task state, member messaging, safe task takeover, an automatic ready-task scheduler, optional quality-gate tasks, and a live activity panel. It is more structured than simply asking one model to role-play several experts in a single response.

Install and verify

The source and curated registry both declare the npm package shown below for the web profile. After the package manager finishes, inspect the composed profile, restart DSH Web, and confirm that the AgentTeams bundle and UI are present before using it on important work.

The copy button records only a consent-gated website interaction. DSH Radar cannot see whether the native package installation succeeds.

Operational limits to understand

• One captain leads one active team at a time.
• Persisted state is serialized inside one DSH process; separate processes do not coordinate edits to the same team.
• Models can finish work without performing the expected task-state update, so the activity panel may need human review.
• Member routes and reasoning effort still consume the configured model providers; the plugin is coordination infrastructure, not free compute.