DeepSeek sketches a Super-agent, and AFK AI's Comma already ships the router-plus-workers design

A QbitAI report published on November 7 pairs two developments: DeepSeek's Harness team recently published an essay outlining a next-generation "Super-agent" architecture, and AFK AI, a Silicon Valley team of Chinese engineers, says its product Comma had already implemented a similar design, with the code released as open source.

The DeepSeek essay starts from a concrete bottleneck: an agent's concurrency can scale almost without limit, while human attention cannot. The team estimates that a person can keep track of roughly three to five active sessions at once, so beyond that point the human becomes the constraint rather than the orchestrator.

Its proposed Super-agent rests on four properties: continuous, efficient, proactive and persistent. That means memory and context carrying across tasks, multiple workers running in parallel, agents that push work forward on their own, and systems that stay online and repair themselves, with the user keeping one entry point and the top-level decisions.

Comma implements this as a single session with a router and multiple workers. A resident router receives instructions and coordinates the job, while workers execute in isolated contexts and report back. The team sums it up as "one team, one conversation": users no longer spin up a separate session per task or re-paste the same background material.

Cross-environment continuity is the harder half. Comma relies on Salix, its in-house distributed agent runtime, which receives events from third-party connectors and uses programmable loops to keep listening and wake agents when needed. The company says it already connects WeChat, Signal, iMessage and Linear.

Memory is deliberately kept outside the model context so that task detail does not crowd out the main conversation. Comma also logs task outcomes, human edits and acceptance feedback, turning them into training signals for a dedicated small model, and a new version only ships if it beats the previous one in the team's own tests.

Once concurrency stops being the bottleneck, the contest shifts from raw model capability to the harness layer: whoever turns multi-task scheduling, cross-device memory and permission control into a dependable product owns the entry point to personal agents. It is therefore not surprising that DeepSeek and a small startup converged on the same conclusion at roughly the same time.

The open questions are specific: stability with dozens of workers running at once, consistency of memory across devices, and whether enterprise permission and information-flow controls survive an audit. Comma supports self-hosting and bring-your-own-key, and can connect to Codex and Claude Code; whether those claims hold at scale still needs independent verification.