DeepSeek open-sourced its agent execution framework, DeepSeek Harness, on 13 August, and per a 36Kr analysis published on 18 September it picked up nearly 100,000 GitHub stars within two days of its public open beta. The framework runs on a tidy equation, 'Model + Harness = Agent', with everything from tool registries to sandboxes built as plugins. The analysis, written from the perspective of enterprise agent firm Hai Zhi Technology, argues the release settles the first big question: the harness is now an independent value layer and, in the way Linux turned operating systems into shared infrastructure, a public chassis. The nervous market reaction, that model makers would drive agent-framework companies out of business, gets the opposite verdict. Not a threat, the piece concludes, but validation and market expansion.
DeepSeek Watch analysis
What happened
DeepSeek Harness is a microkernel framework built on the Cordis meta-framework. Model adapters, tool registries, sandbox environments, session state processors, event dispatchers and interfaces all load as isolated, replaceable extensions, and the execution workflow can be swapped by editing a declarative configuration file. That design, the analysis argues, signals DeepSeek wants the harness to be the chassis standard for the agent era rather than a closed stack.
The piece leans on industry data to explain why this layer matters at all: production agent failure rates of 70% to 95%, with around 60% of failures attributed to data quality, context gaps and weak governance rather than the model itself. The model sets the ceiling. The harness sets the floor.
Key findings
- The analysis is at its best marking out what a general framework cannot do:
- Industry knowledge does not ship with the framework
A harness can ensure an agent executes correctly, not that it executes the right thing; domain rules must be interviewed out of experts and encoded. - The enterprise environment stays hard
Data that cannot leave the domain, heterogeneous systems, approval chains and auditability remain project work, not framework features. - Neutrality is a structural blind spot
A model maker's harness is tuned around its own models, while multi-model routing is now standard on enterprise procurement lists. - Long-horizon memory is unsolved
A CMU, Yale and Amazon survey quoted in the piece calls the harness 'the binding constraint', with context and memory management the thinnest layer of the stack.
Why it matters
If the chassis standardises, value moves up the stack to industry knowledge, trusted data and result accountability. For teams building agents, the floor just became free, and the differentiators are the parts no general framework can preset: your domain rules, your compliance constraints, your accountability chain.
There is a procurement angle too. With multi-model routing now standard, a harness tuned around one vendor's models is a live lock-in question rather than a theoretical one. The piece adds that protocol competition has largely settled, pointing to MCP's major July 2026 revision and the pattern of 'MCP manages tools, A2A manages agents', which pushes the fight to the application layer.
Our read
Weigh the verdict against the author's stake: Hai Zhi sells precisely the layer it declares priceless, so 'the chassis is free, the reins are not' is also a sales pitch. That does not make the boundary analysis wrong, and it is the most useful part of the piece. For anyone building agents, the framework looks worth a trial, and those four gaps are the project plan.
What to watch
- Whether other frontier labs open their own execution frameworks, turning the chassis into a contested standard rather than one lab's gift.
- Adoption beyond stars: production deployments and how deep the plugin ecosystem grows.
- Progress in context and memory management, the layer long-horizon agents need most and currently have least.
- Whether neutrality pressure pushes the harness towards genuine multi-model support, or it stays tuned to DeepSeek's own models.
Discussion spark: When a top model maker hands the agent chassis to everyone for free, does that expand the market for agent builders, or quietly move the moat to a layer the model maker can reach too?
Sources and evidence
not affiliated with or endorsed by DeepSeek