LinkX View | Everything Is a Plugin: Why the Agent Framework Is Gradually Evolving into an Operating System
The stronger the model, the more critical the harness becomes, but its importance manifests in a different way. Patches that make up for the model's shortcomings are disappearing, such as resetting the model after it "loses track" in long tasks. Meanwhile, the components that support agents to work safely and continuously for hours are continuously expanding, including recording what it has done, restricting what it can access, and managing how it connects to new tools. The harness is evolving into something similar to an operating system.
DeepSeek Harness is currently the most open harness from large tech firms, but it is still in its early stage. It is the only large-firm harness that allows external developers to replace the agent loop itself, adopts the MIT open source license, and can also access models from other companies. However, it is currently still in the developer preview version and has not undergone external security audits. Its advantage lies in its room for modification, rather than its out-of-the-box performance.
There are no remaining opportunities for startups in the general-purpose harness space. Patches will be swallowed by more powerful models. The truly durable parts are becoming operating-system-level infrastructure, and model vendors have both the incentive to build it themselves and the ability to tune the harness together with the model. Large tech firms open-sourcing harnesses have also pushed the price of general-purpose harnesses to zero. If any opportunities remain, they are more likely to appear in vertical scenarios built on top of the harness, such as tools, data and workflows for specific industries, rather than in the harness itself.
01 A Harness That Can Install Plugins For Itself
On November 2, DeepSeek launched Creator Mode for DeepSeek Harness. Users describe the desired plugin in plain language, and the agent will first figure out its own operating environment, then complete the whole process of writing code, installing the plugin and running tests. The official demonstration shows that the resulting products include a design workbench, a 3D model viewer, and a sandbox mini-game.
Most software cannot safely "add" new components to itself during operation. Operating systems do this every day: loading drivers, starting and stopping services, and uninstalling them after use, all without requiring a restart.
According to DeepSeek's disclosure, Harness was launched on August 13. In less than three months, its daily active users reached 1.54 million on November 4, with monthly active users hitting 5.48 million. These users consume about 92 trillion tokens every day, and have created a total of about 35,000 plugins.
Based on the comparable weekly average, daily active users increased by 5.2 times in four weeks; after the release of the V4.1-Flash-1017 model on October 17, the growth curve became significantly steeper.
However, the per capita usage has barely changed. DeepSeek's chart shows that on September 30, there were about 16 trillion tokens consumed per day with 271,000 users, and the per capita usage was close to 60 million tokens, which is consistent with the figure in November. The growth is driven by the influx of new users, rather than increasing usage by existing users.
Our Judgment
Agent Harness is splitting into two parts. The half that patches the model is shrinking, while the half responsible for keeping the agent running safely is growing into an operating system.
The so-called patches refer to operations such as resetting the model when it "loses memory" in long tasks; with each new generation of models, a batch of patches become redundant. The operating system half is indispensable for any long-running agent: a record of what has been done, a set of rules defining what can and cannot be accessed, and a mechanism for installing new tools. Most laboratories have reached a general consensus on how to build this part; the disagreement lies in who has the right to modify it. DeepSeek almost allows anyone to replace any component, and we expect other laboratories to move in this direction within a year. What no one can clarify yet is whether users will connect DeepSeek's system to models from other vendors, or only use it with DeepSeek's own models.
02 The Stronger the Model, the Thinner the Harness?
Modifying only the harness can make the score of the same model fluctuate by 16 to 25 points. However, a large part of the harness is used to cover up the model's shortcomings, and this part becomes invalid as soon as the model improves.
What Is Harness
All major laboratories share a common formula: Agent = Model + Harness. The model is responsible for reasoning and generation; Harness is everything outside the model: the loop that calls the model repeatedly, the tools it can use, how memory is managed, and the permissions that determine what it can do.
If we compare the model to a surgeon, the harness is the operating room: surgical instruments, sterile areas, checklists, and the nurse responsible for counting gauze. The more skilled the doctor is, the fewer reminders they need, but no hospital will remove the sterile area because of that. The term "harness engineering" became popular after an article published by OpenAI in February, which described a product with about a million lines of code, none of which was written by humans. The protective gear will affect the results.
OpenAI stated that by only modifying the harness, the score of GPT-5.6 Sol on ARC-AGI-3 increased from 13.3% to 38.3%, and the token usage was reduced to one sixth of the original. Anthropic's data shows that after letting Opus 4.6 filter the content returned by tools on its own, the BrowseComp accuracy increased from 45.3% to 61.6%.
Harness Functions Will Expire
The most typical example from Anthropic is called "context anxiety". As soon as Claude Sonnet 4.5 senses that it is approaching the context limit, it tends to rush to finish the task, so Anthropic added a context reset function in the harness.
By the time Opus 4.5 was released, this flaw disappeared, and the reset mechanism became a burden. Anthropic's advice to developers is: for every new model, ask which tasks in the harness can be removed.
Some Functions May Be Moved Into the Model
Most harnesses use fixed rules to manage memory, such as generating a summary when the conversation reaches a certain length. In late September, researchers from the University of Washington, MIT and Meta published a paper on Context Language Models, which handed this task over to the model itself.
The approach is to give the model its current context in the form of a file, allowing it to modify it as it wishes. Without additional training, this method applied to Qwen3.6-27B and GPT-5.6 Sol outperforms existing baselines and task-specific customized harnesses in long-horizon tasks. The authors said that the next step is to directly train the harness strategy into the model weights.
However, a careful reading of the paper shows that some components have not been moved. The harness still needs to mirror the context as a file and synchronize every modification back before the next round. The authors also had to specially design a new caching scheme called Suffix Cache Reuse to avoid recalculating the entire context when modifying a single part. Moreover, since the model can modify its own memory, injected malicious instructions may also be written in and retained round after round. The memory strategy has indeed moved into the model, but the memory infrastructure and its security issues remain unchanged.
Two Types of Harness
If we classify all functions of the harness by "why they exist", the pattern will emerge. One type is to make up for the model's shortcomings: context reset, fixed compression rules, scripts for chaining tool calls, and filters for organizing tool outputs. Each generation of models will retire several of these functions.
The other type exists because the agent needs to run for a long time on real systems: persistent event records, permissions that the model can never "convince" to bypass, isolated code execution environments, mechanisms for adding and removing capabilities, and cost control. No matter how much the model improves, it cannot replace these components. On the contrary, the stronger the agent is, the longer and higher-risk the tasks assigned to it will be, and the more these components are needed.
Anthropic's own guidelines are actually divided along this line, although the outside world often interprets it as "slimming down the harness". Two of its three principles ask you to rely more on the model and streamline the harness; the third one clearly draws a boundary: place stable content at the front for caching, upgrade high-risk operations to typed tools, so that the harness can intercept, record and audit them. The Managed Agents runtime launched in the same week is much larger than the original Claude Code loop. Anthropic is streamlining the logic of the harness while expanding its infrastructure.
03 The Remaining Half Is the Operating System
The remaining half of the harness performs four tasks, which are exactly the four tasks that the operating system does for programs on a computer:
▪ Remember what the agent has done (save state)
▪ Decide what the agent can access (execution permissions)
▪ Run risky code in an isolated environment (isolated execution)
▪ Allow new tools to be installed (load capabilities)
OpenAI, Anthropic, Google and DeepSeek have already adopted similar approaches for the first three tasks, and the difference lies in the fourth one: how much content external developers can add to the harness and how much content they can replace. In this regard, DeepSeek is the most permissive.
All of this happened very quickly. In less than ten months, all four laboratories have opened their own harnesses to third-party developers. Anthropic Managed Agents in April was the first to split the agent's memory and permissions into independent services. DeepSeek entered the market in August, the latest of the four, but it has the highest degree of openness.
Consensus: State and Permissions
The two laboratories have reached the same conclusion on where to store the state. In DeepSeek Harness, a session is an append-only typed event log, and the model's history is reconstructed from it, so recovery, forking and playback all read the same record. Anthropic's Managed Agents also maintains the same log, which is stored outside the harness and sandbox.
The harness itself does not store state, and any replica can take over any session through four small interfaces: getEvents(), wake(sessionId), emitEvent() and execute(name, input). OpenAI's App Server uses three levels of nesting: threads, turns and entries, which achieves similar effects.
For Anthropic, this split quickly paid off. Now containers are only started when tasks actually need them, the first token latency of typical requests has decreased by about 60%, and the slowest batch of requests has decreased by more than 90%.
Permission control has also reached a consensus. In DeepSeek Harness, every tool call must go through a policy check: the model only makes requests, and the runtime decides whether to approve them; writing "read-only files" in the prompt is only a suggestion, not a boundary. Anthropic simply does not allow credentials to enter the sandbox, so even if the agent is hijacked, there are no keys to leak. Neither design expects the model to be self-disciplined, which is exactly how operating systems treat programs.
Disagreement: Who Has the Right to Modify the System
The real disagreement lies in extension. Claude Code and Codex allow developers to build on a fixed kernel: skills, hooks, MCP servers and sub-agents are attached to the loop, session storage and model binding, all of which are controlled by the vendor. OpenAI open-sourced the Codex runtime under the Apache-2.0 license in August, but its positioning is to use it as a stable protocol for embedding Codex into other products, rather than allowing users to reassemble its suite. Google uses the same harness to support five products, and has made special optimizations for Gemini.
DeepSeek, on the other hand, has made almost everything replaceable. The model adapter, tools, sessions, sandbox, storage, scheduler, interface, and even the agent loop itself, are all plugins running on a small kernel called Cordis. Plugins discover each other through named services such as ctx.llm and ctx.tools, and which plugins to load is determined by a configuration file.
This is the difference between an "app with an extension store" and an "operating system with drivers".
Why would a laboratory be willing to open up to this extent? Because DeepSeek's business is not on the harness. It sells tokens, and tool calls that have been on sale since November 1: the search and crawl API charges $0.75 per thousand requests, and requests sent from Claude Code and Codex are also charged. Laboratories whose profits come from products rely on closed harness moats; laboratories whose profits come from the underlying layer benefit more from opening up the harness, just like Google gave away Android for free back then to maintain its search business. There is another layer of reason: DeepSeek's best model is about four months behind the closed-source frontier, and an open harness can still retain developers even when the strongest model belongs to others.
The Ecosystem Is Real, But Not Yet a Market
DeepSeek claims that users have created about 35,000 plugins, and about 60% of active users have created or installed at least one. For a product that has been on the market for three months, this is very active. However, Creator Mode makes writing plugins as easy as writing prompts, and many of these 35,000 plugins are small tools that only the author themselves uses.
The value of an app store lies in distribution: one person writes, many people install. DeepSeek has not announced any installation volume, and there is no launch date for the app market. Until then, the 35,000 figure only shows that users like to extend the harness on their own, and cannot prove that third parties can do business on it.
Missing Verification
Operating systems are important because they can run on hardware from many vendors: Windows works equally well on Dell, HP and Lenovo machines. For DeepSeek Harness, the corresponding capability is to access models from various vendors, not just DeepSeek's own. On paper, it can do this: developers can access models from other vendors, and some researchers have tested it with Anthropic's Claude Sonnet 5.
What we cannot see is whether anyone is actually using it this way. DeepSeek only counts people whose requests go through its own paid services as Harness users, and people who run Harness with Claude, GPT or Qwen are not included in the 1.54 million daily active users.
04 A Harness That Can Rewrite Itself
DeepSeek dares to let agents extend the harness on their own, relying on a little-known underlying framework called Cordis. Cordis makes the addition and removal of components safe, but it cannot make every component itself safe.
Why Most Software Cannot Do This
Cordis's paper uses VS Code as a negative example. As of June 9, 2026, 87 of the top 100 extensions of VS Code run some code that cannot be completely uninstalled, and the only way to clear them is to restart the process that hosts all extensions. Only 7 extensions declare which other extensions they depend on.
For agents, the cost of restarting is very high. A long-running session will lose the accumulated context and cache, which is exactly the meaning of the harness's existence.
How Cordis Solves the Problem
This framework provides two guarantees:
Clean Uninstallation: Every modification made by a plugin to the shared system is accompanied by an undo operation, and the undo operations are executed in reverse order. Cordis tracks what each plugin registers, such as event listeners, prompt snippets, and tool definitions, and all of them are cleared as soon as the plugin is uninstalled.
Automatic Dependency Handling: Plugins declare which services they need, they are automatically enabled when the services are available, and automatically disabled if any service disappears, just like drivers are loaded automatically when a device is plugged in.
The paper uses seven formal theorems to support these two guarantees, which is rare in agent software. In practice, this means that a Cordis system that changes while running can be analyzed as if it was assembled once and in order.
What Creator Mode Adds
Creator Mode makes use of these two guarantees. Since anything added by the agent can be completely uninstalled, it makes sense to let it try boldly. The agent writes the plugin, installs it, tests it, and if it fails, the system rolls it back.
Context Language Models handed memory management over to the model, and Creator Mode lets the model extend the harness. Both move control to the model, just from two opposite directions.
Where the Guarantees End
Undo is assumed and not verified: The framework does not check whether the plugin's undo operation truly