Skip to content
MaungaAI news, explained

Tools

DeepSeek's agent that rarely interrupts has only been tested in simulation

DeepSeek has openly released a prototype agent that writes code on its own and only gets in touch when it needs to. The company says it finishes sooner than a regular agent, but it has only tested it with fake messages.

2 min read3 sourcesTranslated with AI from the Spanish original

Share: WhatsApp · Email

AI-generated illustration for Maunga. Not a photograph.

The essentials

  • DeepSeek has released dsh-superagent, an agent that writes code in a GitHub repository, under an MIT license: anyone can use and modify it (MarkTechPost, GitHub).
  • Claimed by one partyDeepSeek itself warns that it is a research prototype, not an official product, and that it may have bugs (GitHub).
  • Claimed by one partyAccording to DeepSeek, in a simulation 95% of tasks were finished in under 15.7 minutes, compared with 29.5 with a conventional agent (MarkTechPost).
  • Claimed by one partyAccording to MarkTechPost, by default it only notes down what it would do on its own initiative; acting without being asked has to be switched on (MarkTechPost).

Why it matters

DeepSeek is betting that the agent can hand out many tasks at once and only interrupt when needed, instead of someone watching every conversation. If this holds up outside the simulation, one person could delegate quite a lot more work. For now there is no data on the quality of the code it produces.

The details

The agent is called dsh-superagent and works on a single GitHub repository. When an authorized person asks it for something, on GitHub or, optionally, in the Feishu chat, it opens a work ticket. It handles the ticket in its own session and folder and replies in the original conversation, often with a pull request, which is a proposed change to the code (GitHub). It confirms once, stays quiet while it works and only writes again with a result, a question or a failure (MarkTechPost).

According to DeepSeek's report, cited by MarkTechPost based on the outlet ifeng, one person can only keep a close eye on 3 to 5 sessions at a time, and that slows things down more than the model's power does (MarkTechPost). The tests used made-up messages and simulated tools, and did not measure whether the code it writes is good (MarkTechPost). Also, the released version is not identical to the one the team uses internally (GitHub).

What we don't know

  • Whether the code it writes is good: the tests did not measure that.
  • Whether it works the same way with real messages and teams, not simulated ones.
  • No one outside DeepSeek has checked its figures, which also come secondhand.

Sources

This story is based on 3 sources. In the text, each fact names in brackets who reports it.

Was this story useful? Yes Not really

Spanish original

The thread: DeepSeek

  1. DeepSeek's agent that rarely interrupts has only been tested in simulation · you are here
  2. US accuses six Chinese companies of copying its AI and proposes throwing them off
  3. DeepSeek releases its AI engine, and Latent Space estimates an 80% margin
  4. DeepSeek says its cheap model beats Claude in four tests and loses in five