Digital employee platform StaffDeck goes open source: a cure for AI agents’ workplace syndrome
Over the past two years, every business owner who has seen the potential of AI agents has tried to bring them into existing operations, hoping this group of 24/7 "digital elites" could help cut costs and boost efficiency in enterprise management — simplifying cumbersome processes, answering customers' trickiest questions, and more.
But reality often looks like this: a flurry of impressive moves, and then every deployment is full of bugs.
Many AI agents that can do anything in a demo deck expose a crushing "workplace infant" weakness the moment they enter real business:
Tangled processes: ask an agent to follow a business process and it drifts off course — it should verify the standard before submitting, but submits directly; it should hand off to a human, but confidently gives a wrong answer.
Unsupported answers: decisions (such as "this can be reimbursed") look brisk and confident, but when asked for the basis — which company document, which rule — it simply makes things up.
Same mistake tomorrow: after configuration and release, the error it made on day one is still being made three months later; thousands of conversations' worth of edge cases and user feedback sit there, yet no mechanism lets it learn from experience...
This unreliability is not something a stronger underlying model can fix.
Most agents on the market are still built on chatbot logic — they were designed to chat and entertain, not to knuckle down and work: no process state machine, no knowledge provenance lock, and no self-iteration loop...
As a result, they can chat with you about anything under the sun, but they cannot reliably process an expense report for you within a complex enterprise workflow.
Yet what enterprises want is not an electronic pet that can chat, but a "digital employee" that can shoulder responsibility, perform duties, and accumulate experience.
To solve these problems, ModelBest, together with the Northeastern University–ModelBest Joint Laboratory for Data Intelligence, Tsinghua University's THUNLP, OpenBMB, and AI9Stars, has officially open-sourced StaffDeck, a platform for building and managing digital employees end to end.
StaffDeck does more than let AI take over repetitive communication and standardized tasks; it tries to turn the knowledge, methods, and processes scattered across people, documents, and business systems into organizational capability that is maintainable, reusable, and continuously improvable.
What it answers is not just a technology upgrade, but a revolutionary question: how do enterprises survive in the age of AI agents?
➤ GitHub repository
🔗 https://github.com/OpenBMB/StaffDeck
StaffDeck: redefining the "digital employee" in the agent era
StaffDeck's design premise is to treat every AI agent as a formal employee of the company, not a "problem-solving machine" waiting to be woken up.
In StaffDeck's world, every agent has its own name, position, employee ID, capability boundaries, work records, and even "performance data." Like any ordinary staff member, it must follow SOPs (standard operating procedures), rely on the enterprise knowledge base to support decisions, and accept feedback and correct its behavior in daily work.
On the other side, managers can review the SOPs, knowledge, and tools an agent has mastered, and continue adjusting its capabilities based on real work feedback.
In this way, StaffDeck helps enterprises upgrade and evolve their organizations in the agent era along three dimensions:
From "tool" to "employee": AI is no longer a tool passively waiting for questions, but a "digital stand-in" that has claimed a position and proactively takes on specific business flows.
From "static release" to "continuous operations": launching an agent is no longer a one-off deal; real conversations, negative user feedback, and business edge cases can all inform subsequent adjustments to the agent's knowledge and processes.
From "individual experience" to "organizational asset": employees who know the business can package policy knowledge, judgment criteria, and handling processes into a digital employee avatar, publish it to the employee plaza, and make it available to other members of the organization — turning it into part of the company's assets.
The StaffDeck digital employee plaza
Three core moves to cure AI agents' workplace syndrome
To turn a blustering agent into a reliable digital employee, StaffDeck builds three foundational mechanisms into the AI agent's underlying design, transforming it from a freewheeling workplace slacker into a dedicated model worker.
Process with discipline, flexibility under change (process skills)
In the past, the industry built agents either with "pure skill integration," which is flexible but prone to step confusion in complex processes, or with "pure workflows," which have clear execution paths but struggle to handle tasks outside the flow.
StaffDeck pioneered state-machine-driven process skills, bringing SOPs into digital employees as skills and managing workflows with state machines — each SOP corresponds to a deterministic state machine. Digital employees can switch between multiple SOPs and retain context during the switch, combining principle with flexibility:
Generate an SOP from one sentence: business staff only need to describe a process in natural language, such as "collect invoices first; if over the limit, hand off to a human; if not, submit directly." Within seconds, the system "distills" it into a visual structured flowchart with nodes for information collection, conditional branches, tool calls, and human handoff, which is published after visual confirmation.
Chain multiple SOPs in one conversation: when a user raises multiple tasks in the same request, the digital employee can execute the corresponding processes in sequence and reuse information already obtained.
For example, when an employee says, "File my travel expenses and also check how much of my monthly allowance is left," the digital employee invokes two processes in sequence and "remembers" all the context from the first one, so the user need not repeat any input.
Interrupt midway, resume automatically: when a temporary issue arises during execution, the system can pause the current process, save existing information, and return to the original node to continue after the issue is resolved.
While filling out an expense form, if the employee suddenly asks, "Do client entertainment meal costs count as travel expenses?", the digital employee pauses the current process, looks up the answer in the knowledge base, replies with the source attached, and then proactively brings the user back to the interrupted step: "Let's continue — you hadn't finished entering the invoice amount."
List of process skills
Visual SOP editor
Knowledge with provenance, answers with discipline (OKF structured knowledge base)
The biggest bug in traditional knowledge bases is that they only chunk and retrieve documents, while the AI cannot tell whether a passage is a hard business rule or a historical reference case.
StaffDeck is the first to introduce the Open Knowledge Format (OKF), which grades enterprise knowledge assets. The grading result is as follows:
With knowledge bucket management and retrieval debugging tools, administrators can clearly see in the backend which file and which section each answer hit, and what the relevance score was — ensuring every AI agent answer is traceable and has a source.
Learn the lesson: fix today's mistake tomorrow (feedback iteration loop)
Many agents remain toys because they are static the moment they go live — they do not evolve from mistakes, and the error they made on day one is still there three months later, unable to grow with the company's expectations.
StaffDeck establishes an evolution loop of "full-link tracing + user feedback + human fallback" to solve this.
Every sentence a digital employee says and every API it calls is recorded in a structured way. When a user leaves a negative rating, the system automatically categorizes it (for example, "tool call timeout") and directly points administrators to the direction for modifying the SOP.
Even more elegant is the human fallback mechanism.
When a gray area not covered by policy arises (such as "the exchange-rate conversion basis for cross-border travel"), the AI agent never forces an answer; it preserves the current task state and asks the business creator for help. After a human resolves it, the creator can conveniently add that experience to the knowledge base or optimize the SOP. The next time a similar question arises, it is already an old hand with the answer.
Digital employee work records
Conversation details and task results
The workplace art behind a single sentence
Talk is cheap. After StaffDeck's "workplace training," how does an AI agent perform when it enters a real enterprise business process?
Take financial reimbursement as an example.
In the past, finance teams faced countless repetitive questions every day: "What is the travel standard?", "How do I attach invoices?", "How much allowance is left?" Now they only need a few minutes to write a plain-language paragraph for the "Finance Reimbursement Steward":
When a user applies for travel reimbursement, first collect the reason, amount, and itinerary, then check whether it exceeds the travel allowance standard; if not, continue collecting invoice information and submit the reimbursement form; if it does, forward it to the finance lead for approval.
StaffDeck automatically translates this paragraph into a precise flowchart:
This article includes video content. Please view it in the original WeChat post.
When an employee submits a request:
I traveled to Shanghai on business last week. Please file my travel expenses and also check how much of this month's reimbursement allowance is left.
The reimbursement steward (AI agent) then automatically executes:
It identifies the two tasks, enters the reimbursement process first, and deftly collects the itinerary and amount —
During this, when the employee suddenly cuts in with a question ("Wait, what is the meal allowance for Shanghai?"), it pauses its work, does a precise search in the OKF knowledge base, and tells the employee with a file link and source attached;
Then it automatically takes the employee by the hand back to the interrupted form, continues filling it in, and finally calls the enterprise API to smoothly look up the remaining allowance —
When it hits a blind spot not covered by policy, the agent does not force a perfunctory answer either; it politely packages the whole context and presents it to the finance supervisor. After human approval, the new rule becomes a permanent memory in its mind.
Predictable execution, verifiable grounds, identifiable boundaries, traceable problems — this is the workplace professionalism StaffDeck wants to give every agent digital employee.
Committed to open source: helping enterprises move toward intelligence
Around the creation and operation of digital employees, StaffDeck also embeds enterprise-grade capabilities that support:
Fully private deployment: beyond installing from code, the project provides one-click installers for various operating systems. For institutions with extremely strict data compliance requirements, such as state-owned enterprises and financial institutions, it can run instantly in a local environment — data never leaves the premises, and security is guaranteed.
Scheduled proactive task execution: StaffDeck supports digital employees executing tasks on a daily, weekly, monthly, or one-off schedule and pushing results. In the morning, a digital employee can automatically pull reports, run compliance screening, and push the results to your work group before the meeting.
MCP integration with enterprise systems: native support for the Model Context Protocol, with easy access to ERP, CRM, and OA systems through HTTP APIs. Checking orders, writing reports, submitting approvals — agents can not only talk but actually act.
Digitalization answers "whether work is recorded"; intelligence must answer "whether work can be reliably taken on."
No matter how smart the model is, without a reliable operating system to constrain, guide, and accumulate it, it remains nothing more than a pretty toy confined to a chat box.
The open-source StaffDeck Preview offers five ready-made digital employees (finance reimbursement, legal compliance, HR services, IT support, and administrative steward), each with a complete SOP and knowledge base.
The code is written, the servers are ready, and the ticket for enterprises to enter the agent era is right here. Every issue you file and every PR you contribute will grow, in this open-source ecosystem, into the future of intelligent enterprises in China.
We look forward to developers using it and building it with us:
➤ GitHub repository
🔗 https://github.com/OpenBMB/StaffDeck