DeepSeek Harness team's agent design draws praise for batching: per-decision inference cost falls to 1.6s
teortaxesTex · x · 2026-11-01
teortaxesTex commented on the DeepSeek Harness team's published team-agent design: a solid but expected design, not yet paper material, not better than Dot and similar efforts—but the repo itself is interesting and he expects interesting results to grow out of it.
The attached figures detail the Router layer's batching: as decision-trigger arrival rate rises from 1 to 40 events/minute, average batch size grows from 1.0 to 2.8 while time per request only rises from 2.5s to 3.9s, so inference time amortized per decision drops from 6.4s to 1.6s. When bursts create a backlog, the Router combines queued events for higher effective throughput, and at current team size there's no need to partition the Router layer prematurely at the cost of global context continuity.
Related event: DeepSeek Harness Unveils Super-agent Design and Open-Source Prototype(9 posts)→
More from coding & agent
- GPT-6.1 Terra ranks #14 on Toloka Arena as pass^5 falls to 59% from ~72% — MParakhin · 2026-11-01
- Researcher notes a surge of emails sent by AI agents on people's behalf — anshulkundaje · 2026-11-01
- Coding agent drives iPhone via mirroring to auto-configure shortcuts syncing data to another agent — gethackteam · 2026-11-01
- Alma open-sources its Relay Server for self-hosted, end-to-end encrypted remote access — dotey · 2026-11-01
- Fully vibe-coded browser CS 1.6 clone adds crates and skins, supports 16 players — TAbrodi · 2026-11-01
- MyCaisse ships a remote MCP server: agents read/write freely, money moves only after human approval — Glittering-Cup9298 · 2026-11-01