AI enhances individual efficiency, but processes are what amplify the power of an entire team. Unionman Technology has deeply embedded Feishu CodeM into its R&D workflow and integrated it into standard protocols, transforming it from a personal productivity tool into a collaborative capability accessible to the entire team.
In chip development, there is almost no room for error; firmware testing is a critical step toward mass production, and mistakes often lead to irreversible consequences.
Under such highly cautious conditions, Unionman Technology successfully broke through efficiency bottlenecks. Instead of treating AI merely as a code accelerator, they embedded Feishu CodeM—the native R&D agent within Feishu—into their process system, converting it into a shared capability asset for the entire team.
These changes have yielded quantifiable results: the number of test cases handled in parallel per person per day has doubled; script writing time has been reduced by approximately 70% compared to manual operations; a 10-person team saved roughly 50 cumulative work hours across 10 projects.
More importantly, expert judgment is no longer confined to individual memory. Through one-time configuration and end-to-end flow, senior expertise has been固化 (solidified) into procedural standards, becoming shared R&D infrastructure for the team.
Chip Development Is a "Slow" Industry
Error tolerance in chip development is extremely low. From design to mass production, the cycle typically spans years—and is largely irreversible. Once errors enter the tape-out phase, they are often unfixable. Therefore, the engineer’s core task isn’t writing code quickly, but identifying and verifying every potential risk in advance.
In chip testing, risk identification heavily relies on experience: boundary conditions must be covered, abnormal scenarios cannot be ignored—such judgments usually reside only in the minds of a few senior engineers. This explains why, while the industry broadly embraces “AI coding,” chip engineers remain more cautious—they worry that AI might miss a critical anomaly, which only surfaces after tape-out.
As a controller chip manufacturer, Unionman Technology faces the same challenge. What they needed wasn’t a faster coding tool, but a reliable partner trustworthy at the most critical stages. After evaluation, Unionman decided to introduce Feishu CodeM into their R&D process, starting with firmware testing—the area with the highest stability requirements—to redefine AI’s role within the R&D chain.
The Obstacle to Testing Isn't Difficulty
Prior to adopting Feishu CodeM, the testing team’s work was fragmented across tools. Requirements were recorded in Feishu Projects, but scripting required switching to other development environments. After using AI to generate code, results had to be manually copied back. Significant time was wasted switching platforms and repeatedly copying and pasting.
A deeper issue lay in scattered knowledge. AI performance depends on prompt quality, yet truly crucial expertise—such as key boundary coverage points or timing window controls—is often held only by a few veteran employees. Individual experience couldn’t be institutionalized into team-wide standards, resulting in inconsistent script quality and unstable levels of exception handling and log recording.
Additionally, over a year-long development cycle involving dozens of handoffs, original requirement context gradually diluted during transmission, accumulating risks of missed tests.
Tool fragmentation and knowledge silos ultimately translate into immense pressure on quality assurance.
Embedding Standards Into the Process
The daily work of Unionman’s firmware test engineers involves converting new requirements into Python scripts that can run on devices and automatically generate reports. Due to complex parameters and subjective reliance on experience for exception handling and log formatting, outputs for the same task vary significantly.
Most teams manually invoke AI to generate code and then copy it back into systems—an approach that still treats AI essentially as an advanced code generator, with output quality dependent on human oversight.
Unionman chose a different path: leveraging Feishu CodeM’s deep understanding of R&D logic and collaboration workflows, they directly embedded it into automated testing nodes within Feishu Projects. This created a complete, flowing pipeline—from requirement input, script generation, execution, report output, to manual review.
The key shift is moving quality requirements from post-hoc inspection to pre-defined rules. During requirement input, elements like device information, APK paths, and loop counts are fixed into templates to avoid redundant entry. During script generation, structural frameworks, exception handling mechanisms, and report formats are pre-configured into node rules. The resulting scripts consistently include configuration management, ADB command calls, execution steps, result parsing, and report output—with built-in exception handling—ensuring uniformity regardless of who initiates the task.
Today, Feishu CodeM runs stably in Unionman’s script generation phase. Output configuration files, Excel reports, and exception handling mechanisms are actual deliverables that flow through the system—not mere demo examples. With this quality baseline established, senior engineers no longer need to line-by-line check formatting, freeing them to focus on higher-value human judgment.
Defining a "Release Process" for the Agent
Integrating an intelligent agent (Agent) into the workflow is just the beginning. What truly makes the team willing to entrust it is the built-in constraint mechanism of Feishu CodeM.
This mechanism ensures the Agent strictly follows procedures: during the "technical solution research" stage, file-writing tools are disabled to prevent premature coding; code cannot be submitted before completion or until unit tests pass. If boundaries are breached, the system immediately flags an error and clearly indicates the correct next action.
This set of constraints functions like a clear "release process" defined for a new employee: requirement understanding, solution design, coding implementation, self-validation, and delivery release—all tightly linked. No step can proceed without completing the prior one. The final deliverable is necessarily the outcome of having fully completed all prescribed actions.
Beyond rule-based constraints, there's also the significance of "visibility": the Agent shifts from "passively responding to commands" to "automatically orchestrating tasks according to the process." As requirements flow into a given node, Feishu CodeM actively takes over the task at that point. It is no longer a black-box tool on an individual engineer’s local machine, but a controlled, traceable, and fully recorded standard procedure within the workflow.
This is particularly important in the chip industry. Core R&D data must remain within physically isolated networks; leaking even one schematic could result in losses of hundreds of millions. Feishu CodeM operates within an isolated sandbox, driven forward step by step through the process without requiring human intervention. Every operation is auditable and traceable, perfectly meeting the industry’s strict requirements of “data never leaves the network, full-process auditability.”
Experience Becomes Capability—Reliability Enables Trust
What the chip industry lacks most is not tools, but experience. High tape-out costs and zero tolerance for error make concentration of expertise especially pronounced. Critical tasks cannot be easily handed to newcomers; experts are constantly dealing with emergencies, and when personnel change, experience is lost.
The value of Feishu CodeM lies in transforming these “tacit” judgments into configurable, reusable organizational assets. The same set of process norms, prompt templates, and check rules need only be configured once to become shared across the entire team. Senior engineers’ methodologies are no longer personal secrets; new hires can produce scripts meeting quality baselines simply by following process nodes. Adding a new validation item requires updating the process just once, instantly applying it across all projects. Thus, experience transforms from individual cognition into collective team capability.
Zhu Huanran, IT Lead at Unionman Technology, said: “In the past, senior engineers’ judgment existed only in their own minds. When someone left, the experience was gone. Now, we’ve embedded those judgments into our processes via Feishu CodeM, achieving one-time configuration, team-wide reuse. For us, AI is no longer just an individual productivity tool—it has become shared R&D infrastructure for the entire team.”
The prerequisite for such trust is clearly defining the “boundaries of AI reliability.” Unionman established a clear division: tasks that are rule-based and highly repetitive go to Feishu CodeM, but all outputs must undergo human review. Decisions involving core quality remain firmly under human control. AI handles repetitive labor; humans uphold quality judgment.
AI improves individual efficiency; processes amplify the entire team. At the individual level, one person can now handle twice as many parallel test cases per day, and script writing time is reduced by about 70%. The real leverage appears at the team level—one process configured once, reused by all, replicated across projects. A 10-person team saved approximately 50 hours cumulatively across 10 projects. In this zero-failure industry, Unionman has built, through this process, a reliable mechanism that knows how to guard its boundaries.
Reliability has always been harder than intelligence—and far more worthy of trust.
Turning expert experience into R&D infrastructure is precisely the goal Feishu CodeM aims to achieve for more teams. Today, Feishu CodeM Workflow is officially launched!
Through a visual harness, AI task execution becomes more controllable and systematic, enabling teams to transform their R&D expertise into testable, reusable, and traceable workflows.
Scan the QR code below to “reserve a livestream” and explore the new AI capabilities you need.
If you're also interested in application scenarios of Feishu Projects, click the original link to obtain a "customized solution."