DeepSeek Super-agent
The DeepSeek Harness team’s MIT-licensed always-on coding agent, released 31 October 2026 as an open-source prototype; internal instances run persistently and act without being prompted.
Overview
One-line facts, source-labelled.
| Vendor | DeepSeek (Harness team) |
|---|---|
| Form | Always-on coding agent (open-source prototype) |
| Launched | October 31, 2026 |
| ARscore | not scored |
| License | MIT |
| Pricing | free (MIT); DeepSeek V4.1 Flash API at $0.15 in, $0.60 out per 1M tokens off-peak |
| Benchmark target | Claude Code |
| Open source | Yes (MIT) |
Capabilities
What ships, in the team’s own terms (presence, not benchmark performance).
| Capability | Status | Source |
|---|---|---|
| Unprompted defect detection inside group chat | ✓ | Official |
| Background reproduction of reported defects | ✓ | Official |
| Fix proposals without anyone mentioning the bot | ✓ | Official |
| Parallel multi-task execution | ✓ | Press |
| Persistent internal instances (always-on) | ✓ | Official |
| Open weights and self-hosting (MIT prototype) | ✓ | Official |
Model capability (DeepSWE 1.1 pass rate)
The open-weights models it pairs with, charted by coding pass rate.
DeepSeek V4.1 Flash
open weights, vendor figure
DeepSeek V4 Pro 0813
open weights, independently measured
Bar length equals the DeepSWE 1.1 pass rate. † marks a vendor self-claim that AgentRanks has not independently reproduced; unmarked bars come from the independent table. Source labels: Official vendor announcement, Independent AgentRanks measurement, Press third-party coverage.
Setup and configuration requirements
What you need before first use.
| Source | GitHub repository dsh-superagent-prototype (MIT) |
|---|---|
| Runtime | a persistent host for the always-on instances |
| Interface | team group-chat surface; no mention needed to trigger work |
| Model | DeepSeek V4.1 Flash (open weights; self-host or API) |
| Cost | off-peak API $0.15 in, $0.60 out per 1M tokens |
Use cases
Scenario coverage, not benchmark performance.
CI-style defect triage
Spots defects in the chat stream, reproduces them in the background, proposes the fix.
Parallel fix queues
Runs several tasks at once instead of one session at a time.
Team-chat QA
Works unprompted where the team already talks; no orchestration needed.
Self-hosted Claude Code alternative
MIT licence plus open weights keeps the whole loop on your own box.
Press coverage: Global Times reported the parallel multi-task behaviour on 1 November 2026. Source: Press
Download and official links
Outbound, for trial and verification.
| Source code | https://github.com/deepseek-ai/dsh-superagent-prototype |
|---|---|
| Parent harness | https://github.com/deepseek-ai/deepseek-harness |
| API platform | https://platform.deepseek.com/ |
| News coverage | AgentRanks news feed |
Compare further
The parent project and the scored open-source reference.