DeepSeek's Harness team has published a technical essay outlining a more autonomous "Super-agent" system design, while simultaneously shipping the V4.1-Flash-1017 open-weights model and opening its search and fetch capabilities as a paid API.
A new agent architecture
In the essay "Super-agent Design Exploration: Toward More Autonomous Agent Systems," the team argues that mainstream agent designs still bottleneck on human attention: the number of tasks a person can supervise caps the system's scale. The proposed architecture breaks this limit with a dual-layer structure, event-driven execution, asynchronous multi-channel interaction, compute-storage separation and an external Watchdog. By taking over session management and context synchronization at the infrastructure layer, the system lets humans move up from execution details to higher-level decision and acceptance.
Already running in production
An internal Super-agent built on these principles is already in continuous operation. It participates in bug reproduction and fixes, code-review responses and self-iteration: when a teammate flags a potential defect in group chat, the agent can identify the issue and attempt to reproduce and patch it in the background even without being directly mentioned. This marks a shift from reactive assistants toward agents that proactively maintain their own codebase.
V4.1-Flash-1017 and benchmarks
The new V4.1-Flash-1017 model improves across the board versus the prior V4.1 Flash while cutting runtime and output length, reaching a new performance-cost Pareto optimum under several harnesses. On the Artificial Analysis Intelligence Index it scores 45 (up 6 points, tying for second among open-weights models). Terminal-Bench 4.0 rose from 31.2 to 42.3, and DeepSWE reached 77.0, surpassing Opus 5 (74.0) and GPT-5.6 Sol. DeepSeek positioned the release for "build a complete project from scratch" tasks such as game development and 3D modeling, demonstrating a Forbidden City 3D model, a web racing game and a Portal 2 recreation in the Godot engine from minimal interaction rounds.
A paid Search and Fetch API
Alongside the model, DeepSeek opened its self-developed search engine as an API. Search and Fetch endpoints are priced at 5 yuan per 1,000 calls. Starting 1 November 2026 (Beijing time), server-side tool calls to Search/Fetch triggered inside DeepSeek Harness, Claude Code and Codex begin to be billed, ending the previously free usage for those integrations. The move turns a convenience feature into a metered, revenue-generating service and aligns DeepSeek with the broader trend of agent infrastructure monetization.
Key Facts
Architecture: dual-layer, event-driven, async multi-channel, compute-storage separation, external Watchdog.
Internal Super-agent auto-reproduces and fixes bugs and reviews code without being pinged.
V4.1-Flash-1017: AA Intelligence Index 45 (+6); Terminal-Bench 4.0 42.3 (was 31.2); DeepSWE 77.0.
Search/Fetch API at 5 yuan per 1,000 calls; billed from 1 Nov 2026.
Demonstrated end-to-end builds: Forbidden City 3D, web racing game, Portal 2 in Godot.
Sources: Jiemian News, MBAlib, DeepSeek official; threerouter
Top comments (0)