DeepSeek has published a research prototype that picks up coding requests from team conversations, works on them and returns a proposed change for review. The aim is to reduce the supervision an AI assistant needs, but the developers warn that the code is not production-ready or an official DeepSeek product.
The November 1 edition of AI World Today looks at how the prototype works and what businesses should understand before handing software more responsibility. Also in the episode: keeping work and personal ChatGPT accounts separate, editing a whole presentation with Gemini, and AI helping astronomers process radio signals.
Listen and follow AI World Today on Spotify or Apple Podcasts, or find the show in your favourite podcast app.
From a team request to a proposed code change
The dsh-superagent prototype works with one GitHub repository, where a team keeps its software project. It receives requests through GitHub and, optionally, the workplace messaging service Feishu. When it accepts a job, it creates a work item, handles it in a separate session and responds in the original conversation. The result might be a pull request: a proposed code change that colleagues can inspect before accepting.
Consider a broken download button on a website. A chat-only assistant might suggest a fix, leaving someone to apply it, run a test and report back. An agent connected to development tools can carry out more of that sequence. That is the potential time saving: fewer manual handoffs between getting advice and checking whether the repair works.
The prototype’s design principles also tackle a less glamorous problem: notifications. The agent is intended to interrupt people when it has a usable result, needs a decision or cannot finish, rather than producing a constant stream of progress reports.
That is a useful way to judge automation. If an assistant saves ten minutes of typing but creates twenty minutes of supervision, the job has not become easier.
Permission to start is not permission to access everything
DeepSeek’s authorization rules control which requests become work. For tasks inferred from conversations, the default observation mode records what the system would accept without starting those tasks. A team can review those decisions before allowing more proactive activity.
The limits on execution are a separate issue. The execution documentation says workers run commands without approval under the operating-system account running the agent. Separate working directories do not prevent them from accessing other resources that account can reach.
Giving an agent its own folder is not the same as putting it in a secure sandbox. A folder organizes work; a security boundary restricts access. That distinction matters whenever an assistant moves from suggesting actions to performing them.
For example, a business could authorize an assistant to draft a customer response without letting it send the message. Preparing a refund recommendation need not include permission to issue the payment. The authority should match the job, not simply the capabilities of the model.
ChatGPT makes work and personal accounts easier to separate
A smaller change could be useful immediately: ChatGPT’s mobile account switcher lets users remain signed in to two accounts and move between them without repeatedly logging out. Conversations, memory, files and subscriptions remain separate.
The feature is rolling out on iOS and Android. For people using both personal and employer-managed accounts, it removes some friction, but does not change the policies attached to either account. Check which one is active before uploading a work document.
Gemini can revise an entire Google Slides presentation
Google’s October 30 update lets Gemini edit multiple slides or a whole presentation from one request. Users can change visual styles, revise content using reference files or condense several slides into one. The slides remain individually editable, and changes can be reviewed or reversed.
That could help when adapting an internal presentation for a client. The task is then to check whether the revised deck preserves the facts and commitments that matter. The feature launches in English for eligible Workspace and Google AI plans, with an extended rollout that began October 29.
AI agents are helping astronomers process the sky
In an October 30 report, Nvidia describes agents at the Allen Telescope Array that translate natural-language requests into telescope operations and live radio-data analysis. The company reports real-time AI detection of a pulsar, a rotating stellar remnant that emits regular pulses.
It is a signal-processing result, not a discovery of extraterrestrial life. It also brings the episode back to the same practical question: can software turn a request into an action whose result people can check?
Follow AI World Today on Spotify or Apple Podcasts for daily AI news and explanations.
