Felix Laurent
DeepSeek's Harness team has published a design for an agent that never waits for a prompt. It listens to GitHub, team chats, email and CI, decides what deserves work, and opens pull requests on its own. A research prototype is out on GitHub under the MIT license. The team stresses one caveat early: the article describes architecture ideas, not features DeepSeek Harness already ships.
Why one chat session is not enough
The team starts from a human limit. A person can follow only three to five active agent sessions at once. So single-session agents stop scaling as soon as work arrives from many places. Their answer is a "super-agent" that is continuous, productive, proactive and persistent. It is built from five parts, and the first one sets the shape of the rest.
A router that talks, workers that do
The core is a two-level design. A long-lived router handles every human message and every incoming event, but it never executes tasks. Context-isolated workers do the actual work with the full tool set and never talk to people directly. Events that arrive while the router is still thinking are batched into one decision. The team recommends a lighter, cheaper model for the router.
The router does not poll for work, though. It wakes up on events.
Events instead of prompts
Triggers are ordinary work signals: a chat message, an email, a calendar reminder, a review comment, a failed CI run. A fast rule-based filter goes first, and an LLM judges meaning only when the rules can't decide. Once a worker owns a thread (a PR, a chat topic, an email chain), follow-up events go straight to it and skip the router. The agent can also act speculatively: it drafts a reply or opens a PR while the group is still discussing the issue.
That kind of initiative only works if the agent can listen while it acts.
Full-duplex agents and a watchdog
The usual blocking run() loop is replaced by async start() and stop(), so new input reaches the agent while tools are still running. DeepSeek says Harness already contains this mechanism, but it is off by default. Compute, state and environments are split. Stateless cloud containers run the agent loop, and sessions live in a database. Code runs on the user's own machine through a small daemon. A separate watchdog session checks on a schedule whether each worker is still making progress or drifting from its goal.
The team also measured how much this design saves.
What the first numbers show
The tests used synthetic workloads: about 300 messages and 20 work requests per run, five runs per setup. The team calls the results "indicative only".
| setup | p50 / p95 completion, min | cache-read tokens |
|---|---|---|
| super-agent | 12.4 / 15.7 | 23.2M |
| super-agent without pass-through | 12.6 / 15.5 | 31.8M |
| single session + subagent tool, neutral prompt | 20.4 / 29.5 | 6.1M |
| single session + subagent tool, delegation prompt | 13.8 / 29.0 | 10.8M |
The most telling row is the third one. With a neutral prompt, a single session never delegated work in five runs. The super-agent cuts tail latency almost in half, but it spends far more cached tokens. Router batching also helps under load: at about 40 decisions a minute, the time per decision drops from 6.4 to 1.6 seconds.
Those gains still come with open questions.
What is still unsolved
The team lists the hard parts openly. Memory is the long-term bottleneck for scaling the router. Agent-to-agent communication and context boundaries are unsettled. Security is the biggest one. A prompt injection in a chat message could reach a cloud agent loop. Credentials must also be usable without being visible to the model.
DeepSeek says its internal super-agent already runs nonstop, reproduces and fixes issues, reviews code and helps build Harness itself.
The bigger picture
The super-agent is the next step for DeepSeek Harness, the company's open-source agent runtime. Its developer preview launched on Aug 13, 2026, under the "everything is a plugin" motto. Thehype covered that move into the coding agent race. The design cites Anthropic's managed agents work on "decoupling the brain from the hands". DeepSeek says it targets a different scenario: a user's own machine instead of a fully hosted sandbox. Rivals are moving the same way, toward background agents that pick up tickets and open PRs without a live chat.
The public code is narrower than the article. The README calls it a reference prototype, "not an official DeepSeek product". It is a proactive coding agent for one GitHub repository, with optional Feishu, packaged as a Harness bundle. Some designs from the article are not in it at all.
So is DeepSeek's super-agent ready to replace your coding agent? Not yet. What it does give builders is a clear blueprint for an agent that works from events instead of prompts. It also comes with honest numbers on what that costs in tokens. For teams already running Claude Code or Codex in the background, the router/worker split and the watchdog are the parts worth copying now. The full-duplex loop will matter once DeepSeek turns it on by default.
