DeepSeek's Super-agent blueprint vs Comma: single-session design with shared fact lists and eBPF Loops

机器之心 · wechat · 2026-11-04

DeepSeek's "Super-agent design exploration" proposes four traits for next-gen agents — continuous, efficient, proactive, persistent — arguing "sessions should be a system implementation detail, not a user's cognitive burden." Resident personal agent Comma (launched 6 months before the article) delivers a full engineering implementation of "persistence":

Single-session: one resident Router per workspace receives all messages (group chat, DM, client); tasks go to Worker sessions whose results flow back into one continuous record. Overflowed context is compressed but searchable raw history remains readable.
Condensed Workspace: a shared one-line-per-fact list maintained by a fast decision model, Jev. It gates every write as a probabilistic yes/no question, and after each reply checks whether the list is stale or the agent has drifted — thresholds calibrated on human-labeled samples.
Long-term memory & IFC: Comma for Teams applies Information Flow Control, with a Lean-verified kernel checking that every outgoing message's sources are readable by its audience, preventing private-channel leakage.
Loops: agents write C programs compiled to eBPF for waiting on external events; the async-ebpf runtime carries 40k lines of Lean proofs that programs only touch their own memory — the C compiler itself runs as eBPF in the same sandbox.
Cost-based routing: tasks are classified, success measured by human acceptance, and the scheduler picks the cheapest model+effort combo per category with small exploration traffic.

Timeline: Bridge (Comma's predecessor) launched Apr 28; Anthropic's similar ClaudeTag came Jun 23; Comma launched Oct 1, a month before DeepSeek's article.

Original post →

DeepSeek's Super-agent blueprint vs Comma: single-session design with shared fact lists and eBPF Loops

More from coding & agent