DSH Native Agent Report
Connect DSH Native Agent Report to the DeepSeek API
DeepSeekAgent.io tracks projects that package a task-performing agent directly on DeepSeek Harness. The 26 August 2026 snapshot contains 51 repositories: 17 research agents, 16 engineering agents, 13 creative agents and 5 specialized workflow agents.
Browse every entry in the native Agent directory.
What counts as a native DSH Agent
In this report, a project must provide a Harness-native agent composition—normally an agent.cordis.yml with optional preset.yml metadata—and use that composition to perform a defined job. We include complete DSH forks when they add a distinct agent preset.
We exclude:
- generic agents that merely support a DeepSeek model endpoint;
- UI themes and ordinary capability plugins;
- provider bridges, model selectors and session utilities;
- repositories that mention presets but do not ship a usable composition;
- prompt collections with no native DSH packaging.
This strict boundary explains why the count is smaller than plugin registries containing thousands of packages.
Snapshot
| Category | Agents | Share |
|---|---|---|
| Research, data and finance | 17 | 33.3% |
| Engineering, orchestration and security | 16 | 31.4% |
| Writing, design and game creation | 13 | 25.5% |
| Specialized workflows | 5 | 9.8% |
| Total | 51 | 100% |
The directory grew from 35 to 51 verified entries in the 26 August update, a 45.7% increase. This measures discoveries in our reviewed dataset, not the creation rate of the entire ecosystem.
What people are building
Research agents now cover scholarly search, mathematical reasoning, factor investing, A-share analysis, accounting, data analysis, omics, aquaculture breeding, Scopus search and verify-first scientific workflows.
Engineering agents increasingly look like teams rather than personas: PM-led delivery crews, specialist orchestrators, product teams, expert rosters, operations diagnosis, security assessment and penetration testing. Their durable value usually sits in task state, review gates and role-specific tools—not in a longer system prompt.
Creative agents span novel production, textbook writing, Unity uGUI, Godot, Blender shading, RPG Maker, music-score composition and long-form Chinese fiction. Specialized projects reach GIS, hazardous-waste blending, visual annotation, tea tasting, classical-text collation and beekeeping.
Recurring architecture patterns
Preset plus skills
The most common shape is one selectable preset with several scoped skills. The composition decides the tools and prompt sections; skills carry domain workflows and references.
Lead plus specialists
Multi-agent projects keep one lead in the user conversation and delegate bounded work to role agents. Stronger implementations add persistent plans, evidence files, QA records or independent verification.
Deterministic domain core
Financial and scientific agents often pair model reasoning with Python, Rust, DuckDB or purpose-built CLI calculations. The model chooses and explains; deterministic code computes.
Workbench surface
Some agents add a browser panel or editor for the artifact itself: GIS maps, Unity UI, novel drafts, software-company dashboards or music scores. These are more than chat personas, but the user-facing workbench may be implemented by companion plugins.
Limits of the dataset
Presence of a native composition proves packaging, not output quality, security or maintenance. Repository descriptions are author claims. We inspect structure and public documentation but do not run every agent against a common benchmark.
The snapshot is also discovery-bounded: repositories may be private, unindexed, renamed or described without searchable DSH terms. Counts should be cited with the date and inclusion method.
Suggested citation
DeepSeekAgent.io, “51 Native DeepSeek Harness Agents: Ecosystem Report, August 2026,” dataset snapshot dated 26 August 2026, https://deepseekagent.io/guides/deepseek-harness-native-agent-report
FAQ
Are these ordinary DeepSeek-compatible agents?
No. Every included project must ship a native DeepSeek Harness composition or a distinct Harness fork preset.
Does inclusion mean the Agent is safe or endorsed?
No. Inclusion describes structure and purpose, not endorsement, security review or benchmark performance.
Why not count every DSH plugin as an Agent?
A plugin contributes a capability; an Agent preset composes capabilities into the session that performs a job. Many plugins are useful without being agents.