dsh-swarm

When a single agent can no longer hold the full picture of a complex coding task, dsh-swarm splits the work into parallelizable sub-tasks and orchestrates 100+ concurrent agents inside DeepSeek Harness using a BSP (Bulk Synchronous Parallel) model. The default fan-out is 10 squads of 10 workers each, tiered from a single root planner (deepseek-v4-pro) through squad leads down to worker-level agents (deepseek-v4-flash), each layer owning a distinct slice of the responsibility graph.

Five phases—partition, planning, execution, synthesis, and integration—are closed by strict barriers with straggler deadlines, so no agent ever reads the mid-phase output of another agent. Each squad gets its own git worktree plus a tiered region-ownership map enforced by the sandbox policy, giving you hard write isolation. Guardrail permissions auto-approve in-grant actions, escalate out-of-grant calls to the root planner, and route dangerous or over-budget operations straight to you. Per-tier pricing (pro $1.32/M in, flash $0.0826/M in) with a 90% pre-escalation and a hard mission cap keeps spend predictable; because all workers in a squad share one cache-stable prefix, a 100-worker mission lands at roughly $10-30. Every barrier is a checkpoint, so a mission can resume exactly where it left off.

Built for DeepSeek Harness users who want large-scale refactors, multi-module parallel development, or batch code generation handed to an agent swarm—without giving up write isolation, permission audit, or budget control.