DeepSeek design doc: a person can only follow 3-5 agent sessions at once, attention is the real bottleneck
heyshrutimishra · x · 2026-11-02
DeepSeek's agent team published a design article tackling the question every agent startup avoids: how many agents can one person actually run in parallel. Their answer: models are now good enough to finish complex tasks unattended — the constraint has moved to human attention. In their experience a person can follow only 3-5 active sessions at once; past that, every session is theatre.
Key points from the article:
- Mainstream agent designs treat a single session as the unit of work, with sessions isolated from each other; users juggling parallel tasks must switch manually and track state themselves.
- Bolting on CLIs, MCP, or external connectors to single-session agents doesn't fix the fundamentals: synchronous sessions that block long-running work, mixed contexts across tasks, and everything halting when the local device goes offline.
- Their proposed direction is the Super-agent architecture, removing sessions from human oversight entirely so multi-tasking is limited by model capability, not human attention.
The author frames this as reframing most of what is currently sold as agent infrastructure, since babysitting-per-session products cap at a human ceiling regardless of model quality. The piece is explicitly an architectural exploration, not a description of existing DeepSeek Harness features.
More from AGI Musings
- Carl Frey: 85% in China see AI as beneficial vs 38% in US — resistance maps to expertise costs — carlbfrey · 2026-11-02
- After a painful rebuild, fejau concedes: build on Hermes and you own your AI for life — Teknium · 2026-11-02
- 'Meatwashing': manually rewriting AI output to hide its origin, and why it's a double waste — JeremyNguyenPhD · 2026-11-02
- AI safety researcher: AI companies entirely lack this kind of effective human oversight — davidmanheim · 2026-11-02
- Long-read: AI's real value comes after the efficiency era—infra glut, agent tooling boom — yangyi · 2026-11-02
- vikhyat predicts alternating compute oversupply and undersupply on the road to AGI — TheZachMueller · 2026-11-02