Has the “little crayfish” gone cold?
OpenClaw, which suddenly became popular across China in early 2026, seems to have cooled down already.
Some friends around me who used to stay up all night modifying OpenClaw have quietly put it aside. This project, once said to have exploded on GitHub at a speed rarely seen before, now has almost fallen to freezing point in terms of attention.
Maybe OpenClaw has already finished its historical mission.
It came like a hurricane and, for the first time, really pushed the Agent paradigm into the public view. It made many people realize one thing: AI is not only a chatbot that talks with you. It can also take the initiative and help you get work done.
But OpenClaw did not become Linux.
It ran too fast. Before it had enough time to build deep code barriers and a strong ecosystem, it was already pushed onto the beach.
So, for people who really need Agents to improve productivity, what new tool are they using now?
Recently, an intelligent agent operating system called PilotDeck quietly went online in the community. It was jointly developed and open-sourced by Tsinghua University THUNLP Lab, ModelBest, OpenBMB, and AI9stars.
If OpenClaw is a geek-romantic “big toy,” then this “Tsinghua top student” is more like an agent collaboration cockpit made for real productivity.
PilotDeck Agent OS: From Milk Tea Game to Data Dashboard
What makes PilotDeck different from those early Agents that only tasted the surface?
Let’s look at a few very direct tests.
We opened two WorkSpaces at the same time. One was used to make a game, and the other was used for data visualization.
Then we let them run together, just to see whether it could hold up.
Building a Milk Tea Shop Simulation Game
In the first WorkSpace, we entered one sentence:
“Make a small simulation management game for a milk tea shop. It should have purchasing, pricing, and queuing systems. Customers will decide whether to buy based on price and reputation.”
After the prompt was entered, PilotDeck generated a very detailed plan for the milk tea shop simulation game.
It broke down the core loop of the game design, designed a product line with 5 types of milk tea, and also created the purchasing system, pricing system, customer and queue system, financial system, and so on.
In the technical part, it had already planned a fresh card-style UI layout in advance, and wrote the key JS modules and implementation steps.
In the end, a milk tea shop game could be played online.
Creating an Interactive AI Funding Dashboard
The second WorkSpace went in a totally different direction.
“Here is a group of global AI company financing data. Help me build an interactive data visualization dashboard. It should have animation effects, and mouse hover should show details.”
In this task, PilotDeck used four charts to show the top 10 companies by total financing amount, the financing share of North America / Europe / Asia, and the distribution of general AI / enterprise AI / generative AI tracks.
The final visualization dashboard clearly showed the financing data of AI companies in different regions.
The two tasks ran at the same time.
One was writing game logic.
The other was drawing charts.
They did not interfere with each other.
Making a Programmer Personality Test
After that, we added another purely fun task.
“Make a programmer personality test with 10 questions. It should test what kind of programmer personality you are, and include a result page and a share card.”
PilotDeck generated 10 multiple-choice questions that were quite close to real development scenes. It divided the results into 6 types:
Architect 🏛️ / Brick Mover 🧱 / Perfectionist ✨ / Magician 🧙 / Evangelist 📣 / Philosopher 🤔
The visual style used GitHub dark theme and JetBrains Mono monospaced font. It had a strong tech feeling.
After finishing the 10 questions, the final result showed that I was indeed a Brick Mover.
From a management game, to a data dashboard, to a small social app, the span is honestly a bit ridiculous.
But in PilotDeck, each project is an independent WorkSpace. Each one runs its own task.
PilotDeck WorkSpace: Not Just Folder Isolation
After the tasks were finished, we did something more interesting.
We opened the Memory panels of the two projects separately.
In the milk tea shop project, the memory stored game logic, UI style, and gameplay parameters.
In the data dashboard project, the memory stored chart types, color schemes, and data processing logic.
There was not a single piece of memory mixed between the two.
This is probably the most fundamental difference between PilotDeck WorkSpace and many other tools.
Claude Cowork has introduced Projects for project isolation, and Cursor also has Workspace. But their isolation is still more like “folder + rules.”
The memory is hard to see and hard to modify. Skills do not really evolve as they are used more. And it is not very clear which project spent how much cost.
PilotDeck builds a complete “work cabin” for each project. Inside this cabin, there are three layers.
Dedicated File System
Which files belong to this project, and what the AI has generated, are clearly separated.
Dedicated Memory
Project Memory records the project definition and progress.
Collaboration Feedback records your preferences.
All of them can be viewed, edited, and traced back to their sources.
Dedicated Skills
The Skill app store can install skills into the corresponding WorkSpace with one click.
For example, install game-asset-finder into the game-making WorkSpace, and install minimax-pdf into the document-writing WorkSpace.
Other WorkSpaces may be folders plus static rules.
PilotDeck’s WorkSpace is more like a complete living environment for AI.
PilotDeck Token Routing: Lower Cost Without Losing Results
Agent tools have one open secret.
They are really comfortable to use, but when the bill comes out, it can be scary.
Running the strongest model for the whole task burns tokens very fast. Sometimes it feels more expensive than taking a taxi.
A common solution is to switch models manually. Use a cheaper model for simple questions, and switch to an expensive one for complex tasks.
But this switching itself is also annoying.
PilotDeck uses a set of intelligent routing, and its approach is not exactly the same as common solutions on the market.
Routing by Sub-Agent Instead of Request
The most important design decision is here.
Many routing solutions switch models at the request level. Every request is judged separately, and then assigned to a model.
The problem is that frequent model switching can break the KV-cache. It is like changing models means reading the file again every time. Inference efficiency may actually go down.
PilotDeck does routing at the sub-Agent level.
After a complex task is split into several subtasks, a whole sub-Agent is assigned to one model and runs to the end. The context cache inside this sub-Agent stays continuous.
So what it saves is not only token cost, but also the performance loss caused by switching back and forth.
Flexible Routing Rules for Different Tasks
PilotDeck is also more flexible than fixed routing rules like “expensive models do hard tasks, cheap models do simple tasks.”
It supports rules and prompts to adjust the routing strategy.
You can define what type of task should use what model. You can even tell it in natural language:
“Code-related subtasks use Claude Opus, and text processing uses a cheaper model.”
Open the Routing panel, and every session is listed clearly: whether it is judged as complex, simple, or medium, how much it actually cost, and how much it would cost without routing.
For example, in the programmer personality test app, without routing it would cost $10.97. After routing was turned on, the actual cost was only $1.42.
That means it saved $9.55, around 75%.
The research team also tested this at a larger scale.
In the social media scenario, such as Xiaohongshu content generation, routing cost $2.83. Without routing, it cost $12.58. That saved about 70%.
In more complex task scenarios, such as multilingual podcast processing, financial analysis, and code documentation, the combination of main Sonnet 4.6 + sub MiniMax-M2.7 cost $3.15, with a score of 70.6.
Using Sonnet 4.6 alone cost $18.36, with a score of 69.1.
The price was only about one sixth, but the result was even slightly better.
Of course, if you only care about the strongest performance, you can turn off routing and use the strongest model for the whole process.
The choice is still yours.
Local Models for Privacy and Cost Control
PilotDeck’s routing can also connect local models as sub-Agents.
That means sensitive data does not need to leave your own machine.
For some tasks, it can even decide what tools are needed and automatically deploy an on-device model to do the work. For example, in multilingual podcast processing, it may install VoxCPM by itself to generate speech.
The cloud model can be responsible for thinking, and the local model can be responsible for execution.
In this way, cost saving and privacy are both handled.
PilotDeck Memory: Open the AI’s Brain and Edit It
Today, Agent memory is no longer the old kind of complete black box.
But in many tools, what the AI remembered, when it remembered it, and whether it remembered correctly are still not clear enough.
PilotDeck’s WorkSpace gives a different answer.
It is not just opening a folder. It is opening the whole living environment of an agent.
Open the Memory panel, and every memory has a timestamp, source path, and type.
Project Memory records the core definition of the project.
Collaboration Feedback records your delivery preferences.
If it remembers something wrong, you can click in and edit it.
If there is a memory conflict, you can delete the wrong one directly.
There is no need to restart the conversation, and no need to feed your preferences again from the beginning.
Dream Mechanism for Memory Organization
PilotDeck also has a mechanism called Dream.
During idle time, the AI automatically reviews and organizes its own memory in the background.
It works during the day and digests at night.
In the Memory panel, you can see the Memory Dream button and the Rollback Last Dream button.
If Dream organizes something wrong, you can roll it back to the previous state with one click.
The final result of white-box memory is simple:
The more you use the AI, the more it follows your habits.
Your preferences are stored in Feedback Memory. You can see them, and you can adjust them.
It is not relying on AI to guess what you want.
You tell it. It records it clearly. Next time, it follows.
Why PilotDeck Feels Like a Real Productivity Agent Tool
Looking back at the wave of large models, from the early ChatGPT that surprised everyone, to tools like OpenClaw that tested the boundary of Agent capability like a geek toy, the shape of Agents has been changing very fast.
But only when problems like cost, isolation, and memory are solved one by one can Agents really move from “interesting demo” to “usable productivity tool.”
PilotDeck’s independent WorkSpace, controllable white-box memory, and careful intelligent routing strategy make the idea of “one person, one AI army” look closer to reality.
The more important thing is that this “hexagon warrior” is not locked behind an expensive enterprise paywall.
It chose to be fully open-source.
Whether it is the underlying routing logic or the Workspace architecture, the code has already been placed there openly.
So, if you want to feel what it is like to be a hands-off manager, or you want to open the source code and modify a cyber team of your own, PilotDeck is probably worth a serious look.
