DeepSeek publishes Super-agent coding prototype in challenge to Anthropic's Claude Code

DeepSeek released a prototype of a more autonomous coding agent on 31 October 2026, an offering it calls Super-agent, as the Chinese lab seeks to challenge Anthropic's Claude Code. The code was published by DeepSeek's Harness team on GitHub under the name dsh-superagent-prototype with an MIT licence, and the team set out the design in a post on WeChat the same day.

The pitch is unusual. Instead of waiting for a developer to open a chat window and type a request, the agent can follow group chats where engineers discuss bugs and feature work, then propose fixes without being summoned. DeepSeek sums up the ambition in a line repeated in its manual and repository: take as little human attention as possible, and give as much value as possible.

DeepSeek created the Harness team earlier in 2026 to build agent software that can compete with coding services from leading US labs. In this context a harness is the scaffolding wrapped around a large language model that lets an agent plan, call tools, keep state and carry a task through to a pull request. Anthropic and others have built such harnesses to make their models more effective at programming.

The release lands in a crowded and increasingly political market. Chinese and US developers are racing to ship more capable and cheaper AI tools, and DeepSeek has been raising money at a scale that places it among the best funded challengers. It also arrives while DeepSeek faces restrictions in the United States, where civilian federal agencies were told to remove the company's software.

Key Facts

Bloomberg reported on November 1 that DeepSeek released the prototype on 31 October 2026 with the explicit aim of challenging Anthropic PBC's popular Claude Code. The Harness team said in its WeChat post that the agent can proactively monitor group chats among developers for feedback on coding issues and suggest solutions without being summoned. DeepSeek's own documentation is more modest about the status of the work: the GitHub README labels the code a research prototype for reference only and not an official DeepSeek product.

The Next Web reported on November 1 that the project is published as dsh-superagent under the MIT licence. The agent works on a single GitHub repository and can also listen to Feishu, ByteDance's workplace chat app. When an approved person asks it for work, it opens a task and handles it in its own session, then replies in the original thread, for example with a pull request. A web panel lists what needs a human. The published version differs from DeepSeek's internal code, and some ideas in the accompanying design article are directions for future work rather than shipped features.

The architecture is a Router and Worker split. A long running Router receives human input, applies rules and dispatches work, while isolated Worker sessions each run one task in their own directory. explainx.ai reported on November 1 that the Router can run without a Bash tool and may use a lighter model, while Workers mount the full tool set and need the most capable models. Triggers include chat messages, pull request comments, CI results and emails; programmatic rules filter noise first and a model then decides whether to act. A separate Watchdog session wakes on a schedule and judges whether each Worker is making real progress or has drifted.

MarkTechPost reported on November 1 that the prototype recorded a p95 task completion time of 15.7 minutes in DeepSeek's simulation, against 29.5 minutes for a single session baseline, about 1.9 times faster, with a p50 of 12.4 minutes. Each run carried about 300 messages, of which roughly 20 were real work requests, and each scheme ran five times. The caveat is significant: the test used synthetic messages and simulated tools, and real coding quality was not measured.

The speed comes at a token cost. Super-agent used 23.2 million cache-read tokens in the end-to-end test, against 6.1 million for the single session design, roughly two to four times more. Removing direct event pass-through raised Router decisions by about 30 percent and cache-read tokens by about 37 percent, from 23.23 million to 31.77 million. Proactive work is gated by three accept modes, off, observe and auto, with observe as the default. Setup requires Node.js 22.19 or newer, git 2.32 or newer, Python 3.8 or newer, tmux and GitHub CLI 2.48 or newer. The repository shipped as v0.1.0, a TypeScript bundle released as a single commit, and drew 6,016 stars in three days, including 2,024 on 31 October and 3,992 on 1 November.

Analysis

The most interesting thing about Super-agent is not the model but the unit of work it assumes. Most coding assistants treat the developer as the driver: a person opens a session, types an instruction and waits. DeepSeek's design treats the team conversation as the input stream and the pull request as the output, with a Router deciding which fragments of chatter deserve a Worker. That is a different product shape, and it explains why the team spent its effort on scheduling and noise filtering rather than on benchmark scores.

What this really means is that the competitive ground in coding agents is shifting from raw model quality toward coordination overhead. The simulation numbers support that reading. A single session with a subagent tool available did not use it in five runs under a neutral prompt, so all work queued behind one conversation and was interrupted whenever a new message arrived. With an explicit delegation prompt the baseline improved, reaching 13.8 minutes at p50, but its p95 was still 29.0 minutes. Super-agent, at 12.4 and 15.7 minutes, mainly won by keeping the tail short. The tail is where teams feel pain: one stalled task costs more than a slightly slower average.

The cost side of that trade is real. Using roughly two to four times the cache-read tokens to buy steadier completion is defensible for a team whose engineers cost far more per minute than inference does, but it is a poor fit for budget constrained users or for repositories where most requests are trivial. Any team adopting it should treat the 1.9x headline as a hypothesis to test on its own traffic, not as a specification.

There is a governance story too. An agent that can read a whole team's chat and act on its own is an attention hazard as much as an attention saver, which is why DeepSeek gates proactive work behind off, observe and auto modes and defaults to observe. The Watchdog, the prefilter and the requirement that only an approved person can open a ticket answer the obvious objection: what stops the agent from becoming another noisy participant? The answer here is process discipline, not a capability breakthrough, and that is probably the right order of priorities.

Why It Matters

For DeepSeek, the release is a positioning move as much as a technical one. The company is competing with Anthropic and OpenAI for the emerging market of agents that automate work for developers and business professionals. Bloomberg reported that the latest round brought in about 100 billion yuan, or S$19.2 billion, ahead of a planned initial public offering, while The Next Web cited a figure of about $15 billion. Either way, the war chest is meant to fund exactly this kind of product push.

For buyers, the MIT licence and the public repository lower the cost of experimentation to almost nothing, but they do not remove the integration burden. The stack requires a recent Node.js runtime, git, Python, tmux and the GitHub CLI, and the setup assumes the user already has a coding agent to read the instructions. That recursion tells you who the intended early adopter is: a team already comfortable with agent tooling.

For the wider market, the more consequential detail is the Feishu integration. An agent that listens to a workplace chat platform alongside GitHub moves the assistant out of the integrated development environment and into the place where work is actually discussed. If that pattern spreads, the harness becomes the product and the model becomes a component, which is a harder position for any single vendor to defend.

Next Up

The immediate test is whether outside teams reproduce the scheduling gains on real repositories with real code. DeepSeek's materials say the public code is not identical to the internal version and that some design ideas remain future directions, so the gap between the article and the repository is the first thing independent evaluators will probe. Watch also for the accept modes to be exercised in the wild, since observe is the default and auto is where the risk sits.

Beyond that, the roadmap implied by the design post is broader trigger coverage and a smarter Watchdog. The unresolved question is regulatory rather than technical: the same autonomy that makes Super-agent useful is what makes it awkward in markets where the company already faces restrictions. How DeepSeek handles that tension will shape whether this prototype becomes a product line or stays a research artefact.