Open-source AI agent frameworks are growing fast, and the latest number comes from China. DIGITIMES reports that DeepSeek says its open-source agent framework, DeepSeek Harness, had 1.54 million daily active users as of November 4, 2026, more than quadrupling in under three months. The same report says the plugin ecosystem is expanding and an official plugin marketplace is reportedly planned. The article is paywalled, and we have not seen the framework’s documentation or source, so we treat everything beyond those headline figures as unknown.
The number matters less for what it says about DeepSeek than for what it says about the market. Developers are picking open frameworks over proprietary agent platforms, and they are doing it at scale.
What we know about DeepSeek Harness, and what we don’t
Known, per DIGITIMES: the framework is open source, DeepSeek reports 1.54 million daily active users, growth has been more than four times in under three months, and a plugin marketplace is reportedly in the works.
Not known: the license, the supported languages, how “daily active user” is counted, whether the figure includes DeepSeek’s own hosted products, and what the plugin API looks like. Self-reported usage numbers deserve caution until a third party checks them. We will update this piece when we can read the primary material.
Separately, release trackers list a DeepSeek model called deepseek-v4-flash-0731 as published on November 5. We have not tested it and cannot say how it relates to Harness, if at all. A cheap model paired with a free framework would make sense, but that is inference, not reporting.
The open-source framework landscape in November 2026
Harness arrives in a crowded field. Shakudo’s November 2026 roundup is a reasonable map of who is who. The frameworks that come up most often:
| Framework | Maintainer | Design | Notable in 2026 |
|---|---|---|---|
| LangGraph | LangChain | Graph of nodes and edges, built-in checkpointing | Still the default for stateful, long-running agents |
| CrewAI | CrewAI | Role-based “crews” of agents | Fastest path to a working demo |
| Microsoft Agent Framework | Microsoft | Merged Semantic Kernel and AutoGen | v1.0 reached general availability on April 2, 2026 |
| Google ADK 2.0 | Graph-based execution engine | Announced at I/O 2026; reportedly added multi-language support and security guidance this week | |
| Claude Agent SDK | Anthropic | Opinionated loop with sensible defaults | Tied to Claude models |
| Auto Ontology | NVIDIA | Python, graph-based semantic reasoning | Reported around November 6; early days |
| DeepSeek Harness | DeepSeek | Unknown; plugin-based per reports | 1.54M DAU claimed, Nov 4 |
Two patterns stand out. First, graphs won. LangGraph, Google ADK 2.0, and NVIDIA’s Auto Ontology all model an agent as a graph of steps, because a graph is easy to checkpoint, replay, and draw. Second, consolidation is real. Microsoft folding Semantic Kernel and AutoGen into one framework is the clearest sign that the experimental phase is ending.
We compared three of these hands-on in our framework comparison, building the same support-triage agent in each.
Why developers are choosing open frameworks
Cost
An agent framework is mostly plumbing: a loop, a tool registry, some state. Paying a per-seat or per-run fee for plumbing is hard to justify when the model calls are already the expensive part. Open frameworks move that cost to zero and let you spend the budget on tokens.
Control
When an agent misbehaves in production, you need to read the orchestration code, not file a support ticket. Open source lets you step through the loop, patch the retry logic, and pin the version. Closed platforms give you a dashboard.
Model-agnostic by default
Most open frameworks treat the model as a swappable adapter. That is leverage. If a cheaper or better model ships next month, you change one line. Frameworks tied to a single provider, including Anthropic’s Claude Agent SDK, trade that flexibility for tighter integration. That is a fair trade for some teams and a dealbreaker for others.
Self-hosting for privacy
Regulated teams cannot send customer data to a third-party orchestration service. An open framework runs inside your own network, with your own model endpoint, and your logs stay with you. For healthcare, finance, and government buyers, this is often the deciding factor.
What a plugin marketplace means for lock-in
A marketplace sounds like openness, and in one sense it is: more tools, less code to write. But marketplaces also create gravity. Once your agent depends on twenty plugins written against one framework’s API, moving to another framework means rewriting twenty integrations.
This is the dynamic that made npm and the Chrome Web Store so sticky. The framework stays open; the ecosystem becomes the moat. If DeepSeek does ship an official marketplace, watch the license on the plugin API and whether plugins are plain functions with a thin wrapper or something tied deeply to Harness internals.
The practical defense is the same as it has always been: keep your tools as ordinary functions with their own tests, and keep the framework-specific glue thin enough to rewrite in a day.
How to evaluate an agent framework in an afternoon
You can learn most of what you need in three hours. Build a two-tool agent and check five things.
- Checkpointing. Kill the process mid-run. Can you resume from the last step, or does it start over? Long tasks and human approvals both depend on this.
- Tracing. Does every model call and tool call land in a log you can read, with inputs and outputs? If you need a paid add-on to see what the agent did, note that in the cost column.
- Tool schema. How do you declare a tool? A clear schema with types and descriptions means the model calls tools correctly more often and lets you reuse the same tool elsewhere.
- Human-in-the-loop. Is “pause and wait for approval” a first-class step, or a pattern you have to build? LangGraph and Google ADK treat it as built in; several others leave it to you.
- License. Read it. Apache 2.0 and MIT are safe. Anything with a commercial-use clause, a source-available license, or a contributor agreement that assigns rights deserves a conversation with whoever signs your contracts.
If a framework passes all five, run it against your real task for a week before committing. If it fails two, move on.
Our take
DeepSeek’s 1.54 million figure is unverified, but the direction is not in doubt. Open frameworks have become the default way to build agents, and the competition now is about ecosystems, not core loops. That is good for developers in the short term and worth watching in the long term, because ecosystems are where lock-in hides.
Our advice: pick an open framework that passes the five-point check above, keep your tools portable, and do not let a plugin catalog make the decision for you. Start with what an AI agent is for the fundamentals, see our comparison for the trade-offs, browse the tools directory, and get the weekly update in the newsletter.
Questions, answered
What is DeepSeek Harness?
DeepSeek Harness is an open-source framework from DeepSeek for building AI agents. The company says it reached 1.54 million daily active users on November 4, 2026, more than quadrupling in under three months. Full technical details are behind a paywalled report and have not been independently confirmed.
Why do developers prefer open-source agent frameworks?
Cost, control, and portability. An open framework lets you swap models, inspect and patch the orchestration code, and self-host for data privacy, without being tied to a single vendor's pricing or roadmap.
How do I evaluate an AI agent framework quickly?
Check five things: checkpointing and resume, tracing of every model and tool call, a clear tool schema, a built-in way to pause for human approval, and the license. If any of those is missing or vague, keep looking.
