DeepSeek, the Hangzhou-based AI lab, has published the code for Super-Agent, an experimental coding assistant designed to work for a whole software team rather than a single developer. The project is available on GitHub under the permissive MIT licence, and the company describes it as a research prototype, not an official DeepSeek product.
Bloomberg reported that the release is aimed at Anthropic’s Claude Code, one of the best-known AI coding tools. DeepSeek’s team laid out the thinking behind the design in a WeChat post at the weekend.
How Super-Agent works
Most coding assistants wait for a developer to type a prompt. Super-Agent is built to sit where a team already talks about its code. It connects to a single GitHub repository and can optionally listen to Feishu, the workplace chat app owned by ByteDance. When an approved team member asks for something, it opens a work item, handles the job in its own separate session and workspace, and reports back in the original thread, often with a pull request ready for review. A web panel lists the items that need a human decision.
Reports say the agent can also spot coding problems raised in group chats and suggest fixes without being called. DeepSeek frames this cautiously. Its stated design rule is to take up as little human attention as possible while delivering as much value as possible, and it warns that an agent working for a whole team can easily waste that team’s time. The agent is therefore meant to act only when asked, or when the team’s own rules allow it.
The company is open about the limits. The public code differs from the version DeepSeek runs internally, some ideas in its companion design essay have not been built yet, and users are told to review and adapt the code before running it. It handles one repository at a time, and it does not decide who may give it instructions or what it may change; operators set those rules.
The release fits a wider push. DeepSeek set up a team earlier this year to compete with American coding tools and previewed a broader open-source agent app in August. Rivals are heading the same way: GitHub recently showed a Copilot app that runs agent tasks across several repositories in parallel. There is a known catch, though. A study published in January found coding agents performed about 30 per cent worse on average when working as a team than alone. That study did not test Super-Agent.
What it means here
For Bangladesh’s software and IT services firms, many of which build and maintain code for overseas clients, an open, MIT-licensed agent changes the cost calculation. Teams can read the code, run it on their own servers and adapt it to their workflow without paying per-seat fees to a foreign provider. That matters for small studios with tight budgets and for clients who are wary of sending source code to third-party clouds.
There are caveats. Super-Agent is a prototype, its chat integration targets Feishu rather than Slack or Microsoft Teams, and fitting it into a local setup would take real engineering work. Firms with American clients should also weigh the politics: US federal civilian agencies were told last month to remove DeepSeek software, and some clients may ask pointed questions about tools linked to the company.
The bigger signal is the direction of travel. AI coding tools are moving from autocomplete towards agents that pick up tickets and open pull requests on their own. Local firms that learn to supervise, review and audit that kind of work, rather than simply producing more code faster, will be better placed as overseas clients begin to expect it.
