DeepSeek open-sources Super-Agent — and it defaults to watching
DeepSeek's Harness team published two things over the weekend: a design article arguing that the binding constraint on AI agents is no longer the model but the human coordinating them, and a working implementation of its answer. The code is on GitHub as dsh-superagent under the MIT licence, and the repository had passed 5,900 stars by Sunday afternoon, about thirty hours after it appeared.
The framing in the coverage is that this is DeepSeek's shot at Anthropic's Claude Code. Bloomberg's Debby Wu reported the release as an attempt to rival it, and the South China Morning Post's account describes an agent that watches developers' group chats and proposes fixes without being summoned. That is the pitch. The shipped code is a good deal more cautious, and the gap between the two is the most useful thing in the release.
What actually shipped
Super-agent is a bundle for DeepSeek Harness, the company's MIT-licensed agent runtime. Out of the box it watches one GitHub repository: new issues, new pull requests, comments, and the default branch's CI results. Feishu, ByteDance's workplace chat app, is an optional second source. When someone on an approved list asks for work, the agent opens a ticket, does the job in its own session and its own directory, and replies in the original thread, often with a pull request. A web panel shows what is waiting on a person.
The architecture is borrowed from operating systems, not from prompt engineering. Incoming messages become triggers. One routing session decides, per trigger, whether to accept it as new work, attach it to existing work, stop something, or ignore it. Each accepted ticket gets one worker session from a pool capped by a maxWorkers setting. All state lives in a single SQLite file, and replies go through a durable outbox that fires only after the database commit, so a crash mid-turn does not lose the answer or double-send it. The design article adds pieces the prototype lacks: a two-stage filter that lets cheap rules discard most traffic before a model sees it, a watchdog that catches agent loops and drift, and a split between compute, session state and local execution so an agent survives your laptop closing.
One detail is easy to miss. A worker that picks up a pull request can subscribe to that thread directly, so later review comments land in its session without routing through the main coordinator each time. DeepSeek's own measurement of removing that shortcut is the most honest number in the release: router decisions rise about 30% and cached-read tokens go from 23.23 million to 31.77 million, according to 智东西's read of the paper. Routing is not free, and the team priced it.
The default is the story
Every headline says the agent acts without being asked. The authorization document says otherwise, and it is unusually explicit. There are three accept modes for work nobody requested: off, observe, and auto. Observe is the default, and in observe a positive assessment of an inferred task is "recorded as ignore marked 'would accept'" while nothing is accepted. Even in auto, a proactive accept requires at least one matching rule that the team wrote itself. Explicit requests work in every mode.
So the honest description of v0.1 is: it watches, it judges, it writes down what it would have done, and it waits for you to promote it. DeepSeek's own guidance is to run in observe, read the recorded decisions, then switch. That is a sensible way to ship an agent that acts on a team's behalf, and it is also the opposite of the autonomy the launch copy implies.
The README adds a second caveat that journalists skipped. The published code was adapted for release, so it is not identical to DeepSeek's internal version, and some designs in the article are described as future directions that the prototype does not implement. What you can install is a subset of what you can read about.
The numbers, and whose they are
DeepSeek ran one experiment: a simulated collaboration environment of about 300 messages and 20 work requests, comparing Super-agent against a single session with subagent tools, five repeats per configuration. Median task completion came in at 12.4 minutes and the 95th percentile at 15.7 minutes, against 29.5 minutes at the 95th percentile for the conventional setup, as 量子位 QbitAI reports. The win is in the tail, not the median, which is the number that matters if you care whether an agent occasionally strands a task for half an hour.
Treat all of it as a vendor benchmark on a vendor-built environment, with five runs per arm and no independent replication. It is a plausible result, not a demonstrated one. The structural claim underneath is the one that will survive or not: a coordinator that manages context and workers separately should beat a single session that has to do both, because the single session's context is the scarce resource. If that is true, it is true regardless of whose model sits inside.
Why the attention bottleneck is the real argument
The paper's central line, quoted by the SCMP, is that "single-task limits depend on model capabilities, while multitask limits depend on human attention." The team estimates a person can keep track of three to five active agent sessions before the bookkeeping eats the benefit. Anyone who has run parallel coding agents recognizes that ceiling immediately: the work is not writing the prompt, it is remembering which session holds which decision and re-explaining context when you switch.
That reframes what a harness is for. We wrote about this when OpenAI started renting out the machinery behind Codex, and it is now the whole competitive axis: not the model, the session manager around it. DeepSeek shipping its coordinator as MIT code, on top of a harness it also open-sourced, means the coordination layer stops being a moat and becomes a commodity that any lab can fork. Anthropic's advantage in Claude Code is now partly a design advantage, not a weights advantage. DeepSeek's own models are a separate matter, and cheap ones: V4.1-Flash-1017 hit 77% on DeepSWE for 20 cents a task, which is the sort of price that makes a chatty, always-on coordinator economically thinkable in the first place.
What the skeptics have
The strongest counter-evidence predates this release. A January study called CooperBench found that coding agents working as a team did worse than the same agents working alone, with success rates about 30% lower, as The Next Web reported in its writeup. DeepSeek did not test against that benchmark, and its own experiment measures completion time, not whether the code was right. Speed is the easy metric for a coordinator; correctness is the hard one.
The other objection is reliability, and it came from a Chinese practitioner rather than a competitor. Liu Ming, a Beijing AI entrepreneur, told the SCMP that the hard problem in AI software development is "ensuring reliability rather than completing tasks faster." DeepSeek concedes the related risk in its own writeup: an agent wired into chat, code and CI is a prompt-injection surface, where a malicious instruction in any connected service can steer work it was never meant to touch. That is the same failure class we picked apart after the Cursor marketplace backdoor ran for eleven hours, except here the agent has commit rights by design.
Scope limits the damage for now. One repository, one optional chat app, and a default mode that refuses to act. But the moment a team flips to auto and writes permissive rules, the interesting failure is not a bad pull request. It is a coordinator that quietly decides a discussion in a group chat is a work request, opens a ticket, and spends an afternoon on it.
What to watch
Three things. First, whether observe stays the default when DeepSeek folds this into Harness as a product: the default is the whole safety story, and defaults drift. Second, whether anyone publishes a correctness score rather than a latency score, on a task set DeepSeek did not build. Third, whether the harness's plugin model attracts the same fine-tuning crowd that turned MiniMax's video base into an ecosystem, because that, not star count, is what decides whether an architecture becomes a standard.
For now the honest summary is narrow and still notable: DeepSeek has published the cleanest public design for multi-agent coordination anyone has shipped, and shipped a version of it that watches instead of acting. Both halves are the point.
