DeepSeek just gave the open-source world another peek inside its workshop. On October 31, its Harness team open-sourced dsh-superagent-prototype — a research prototype for a new kind of coding agent — under the MIT license. The twist that has developers talking: much of the prototype itself was built by coding agents, in the team's daily development work. It crossed 2,100 GitHub stars on release day.
What makes it a 'Super-agent'?
Most coding agents today work one session at a time. You open a chat, give it a task, wait. Juggle three tasks and you're switching sessions yourself — and, as DeepSeek's team notes, most people can only follow about 3 to 5 active sessions at once before things get messy. The Super-agent design moves session management into the system. A two-layer Router-and-Worker setup sits at the core: the Router talks to humans, keeps the global context and dispatches work, while Workers run concurrently in isolated sessions and report results back instead of replying to users directly. Events from GitHub and Feishu (a workplace chat tool) become managed tasks, each with its own session and directory, so several GitHub issues can move forward in parallel.
Built around human attention, not benchmarks
The design's most interesting idea is its constraint: minimize human attention cost. An agent that fires off twenty notifications, opens four competing draft pull requests and asks five vague questions costs more engineering time than writing the code by hand. So the prototype filters trigger events, asks a single structured question when it genuinely needs a human, and releases its worker slot while waiting. An external watchdog keeps the whole thing running persistently — tasks don't die just because a local machine goes offline.
Why it matters
This is a research prototype, not a finished product — DeepSeek is candid about that. But the direction is clear: agents are graduating from single-task assistants to systems that manage whole workloads. If the Super-agent approach holds up, the bottleneck in software teams shifts from writing code to deciding what to build and reviewing what agents produce. And because it's MIT-licensed, any team can read the design docs, run it, and adapt it — the same open playbook behind DeepSeek's earlier open-source engine, DeepInfer.
The take
The headline writes itself — an agent largely built by agents — but the real story is quieter: DeepSeek is
treating developer attention as the scarce resource, and designing around it. That's a bet worth watching, whether or not this exact prototype becomes a product. Related reading: GitHub's Copilot Fusion and Cursor's production-babysitting bots.
Open source, parallel tasks, and a design that respects your attention — DeepSeek's Super-agent is a small release with a big hint about where coding agents go next.
Comments
Post a Comment