DeepSeek open-sources dsh-superagent: Router-Worker agents ~1.9x faster than single-session
Available_Damage2374 · reddit · 2026-11-05
A Reddit user reports DeepSeek released an open-source research project, dsh-superagent (v0.1.0, MIT), focused on the agent harness rather than model weights (community report, not yet officially confirmed).
Core thesis: the bottleneck for multitasking is no longer model intelligence but human attention — developers can only babysit 3–5 single-session agent windows before losing track.
Architecture:
- Router: listens to GitHub events and team chat, filters noise, tracks state, dispatches tickets, and only pings you to confirm, unblock, or hand over a finished PR.
- Workers: each ticket gets an isolated agent session and directory, running in parallel without polluting each other's context.
Results: in a simulated noisy workload (300 chat/CI messages with 20 real tasks), it hit p95 completion of 15.7 minutes vs 29.5 minutes for a standard single-session agent (1.9x faster). A gated proactive mode (default observe) watches repo/chat streams and logs tasks it would pick up autonomously.
Related event: DeepSeek Open-Sources dsh-Superagent Router-Worker Framework(2 posts)→
More from coding & agent
- LMArena opens Agent Arena for agent evaluation — arena · 2026-11-05
- Running Qwen 27B Q4 at 0.7 tok/s on a 4GB-VRAM laptop as a main coding model — Felix_455-788 · 2026-11-05
- His agent now auto-labels his Google Calendar in the background — steren · 2026-11-05
- The security startups racing to rein in AI agents — pstAsiatech · 2026-11-05
- OpenAI October dev roundup: Decisions API, 12-hour caching, 8x faster Codex tokens — OpenAIDevs · 2026-11-05
- Datadog previews Canvas, a new way to collaborate with AI agents on observability — farooqib · 2026-11-05