- The dated hands-on report describes a persistent router delegating work to workers with separate contexts.
- Comma's earlier product milestones are distinct from November 7 reporting; its learning mechanism is described rather than independently benchmarked.
Hands-on task coordination
QbitAI's November 7 hands-on report compares AFK AI's Comma with DeepSeek's proposed Super-agent architecture, focusing on a persistent router coordinating workers with separate contexts. The comparison addresses a shared problem: users' limited attention when supervising many concurrent agent sessions.
In the reporter's account, Comma delegated three requests—building a cooking game, preparing a weekly review and organising AI model-release information—to different workers from one conversation. The report also describes creating Linear issues through Comma and passing Linear tasks back for execution, with progress returned to the original task.
Memory outside the conversation
The article says Comma stores long-term state outside the model's context and separates the router's main conversation from workers' detailed task contexts. It describes a learning mechanism that uses task outcomes, human edits and acceptance feedback to update specialised small models, with replacements required to beat predecessors in the team's own tests.
These are reported product mechanisms and hands-on observations, not independently reproduced performance benchmarks. The article dates Comma's open-source release to September 30; November 7 marks this comparative explanation and testing report rather than the product's initial release.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01
