DeepSeek open-sources Super-Agent, an always-on AI coding agent aimed at Claude Code

DeepSeek has released Super-Agent, an open-source AI coding agent architecture designed to run continuously in the background rather than inside an interactive terminal session. According to a post on dev.to, the project was published by DeepSeek's Harness team on GitHub under the MIT license, under the repository name dsh-superagent-prototype, and is explicitly positioned as an autonomous alternative to Anthropic's Claude Code.

A router, workers and a watchdog

The architecture is built around a two-tier "router and worker" model. The router is an always-on, lightweight agent that watches event streams such as group chats on Slack and WeChat, GitHub pull requests and CI/CD pipelines. It acts as a triage layer: it prioritises incoming triggers and delegates work, but never executes shell commands itself.

Actual work is handed to workers, isolated execution sessions spun off by the router. Each worker runs a full-capability LLM with a shell toolset and tackles a specific coding task, and multiple workers can operate in parallel. A third component, the watchdog, is a timer-driven monitoring session that tracks active workers so that tasks do not stall or get stuck in infinite loops.

Autonomy instead of terminal sessions

The core contrast with Claude Code is how work gets triggered and where the developer's attention sits. Claude Code, as the dev.to post describes it, is a terminal-first, synchronous tool: the developer opens a repository, types instructions at a CLI prompt and responds to permission requests as the work proceeds. Super-Agent instead runs as a background process or service, listening passively for CI failures, PR comments and bug reports in chat channels. According to the post, it can read ongoing developer conversations, detect bug reports or technical discussions, and offer fixes or code branches before anyone explicitly tags it.

The design also leans into parallelism. Rather than tying a developer to one terminal session at a time, the router instantiates a separate isolated worker for each inbound issue. The dev.to write-up gives the example of five pull requests being patched simultaneously across multiple repositories without keeping five terminal windows open.

Guardrails and the cloud-local split

Because an always-on agent that writes code unprompted is inherently risky, the architecture introduces explicit permission tiers: off, observe and auto. In observe mode, a worker prepares an entire solution, logs the shell actions it would have executed, and opens a GitHub draft pull request or a Slack summary for human review. Nothing is applied to the base branch until a person approves it.

The infrastructure is deliberately decoupled as well. The orchestration loop runs persistently in the cloud or as a hosted daemon, while a lightweight local daemon performs terminal actions on the developer's machine only when needed. Workers are described as ephemeral daemons that spawn when assigned a task, pull down the relevant context, run tests and tear themselves down after submitting results.

Reported performance, with caveats

According to the dev.to post, synthetic benchmark tests published by DeepSeek show Super-Agent completing tasks with a 95th-percentile time of 15.7 minutes, compared with 29.5 minutes for a traditional single-session subagent approach. Those are the project's own synthetic numbers relayed through a single source and have not been independently verified. It is also worth noting that the repository is framed as a research prototype rather than a finished product.

Why it matters

The pitch reframes what AI coding tools are competing on. Instead of arguing about raw model capability, Super-Agent targets human attention, on the argument that developer time, not model intelligence, is the real bottleneck in software development. If a background agent can triage chats, CI failures and PR comments and turn them into reviewable draft fixes, the developer's role shifts from directing a tool step by step to reviewing a queue of proposals.

The licensing matters too. An MIT-licensed orchestration layer lets teams inspect the plumbing, self-host it and adapt it, in contrast with proprietary CLI products like Claude Code. The open question is governance: agents that act before being asked raise real safety and review-process concerns, which is exactly why observe mode and the watchdog exist. As it stands, Super-Agent is less a shipped product than a statement about where agentic developer tooling is heading.