Comma, an always-on personal intelligent agent, adopts a single main session design: each workspace retains only one main session that aggregates all entry points, with the Router sharing context and tasks handed off to independent Workers for execution—thus solving the problem of context forgetting. Comma maintains a concise workspace fact list shared by all Agents through its rapid decision model Jev, determining whether content is worth sharing, whether it has become outdated, or whether a task has deviated. The team edition manages information access permissions for different audiences through an information flow control mechanism, with long-term memory stored partitioned by audience and supporting direct user modification. In addition, Comma supports Agents in writing their own resident monitoring program Loop, compiled to eBPF, to handle external waiting events, and dynamically adjusts model scheduling strategies for different types of tasks based on manual acceptance results and task costs. This design predates the public disclosure of related concepts behind DeepSeek's Super-agent, allowing user-specified requirements to be continuously remembered and carried forward—enabling the agent to work competently over the long term.