dsh-agent-swarm

When a task spans multiple files, requires decomposition and coordination, or involves a completely unfamiliar legacy codebase, a single agent simply cannot deliver—lacking division of labor, cross-validation, and any recovery path once interrupted. Agent Swarm brings a collaborating team of agents into DeepSeek Harness: the primary agent inspects the project, sets the plan and acceptance criteria, and allocates resource budgets; members work in parallel inside their own Git worktrees, exchanging evidence, challenging findings, and handing off partial results; a separate member independently reviews submitted work, and the declared checks are executed against the exact commit.

The sidebar is your window into the whole team: it shows the current operation of every member, elapsed time, and the latest event stream in real time, with pause, resume, and stop controls always at hand. If a task is interrupted, durable state on disk—assignments, evidence chains, and decision logs—lets the team resume from checkpoints. Replacement tasks inherit the original acceptance criteria, downstream dependencies follow automatically, and you never have to re-describe the goal.

It is built for developers working in local Git repositories whose tasks genuinely benefit from division of labor and cross-validation: fixing a cross-module bug, delivering a feature that touches several parts of the codebase, or tracing through unfamiliar code without modifying a single line—all are natural Swarm scenarios.