A source code analysis of MetaGPT's mobile testing module circulated in the tech community this week. The analyst's conclusion is blunt: MetaGPT's famous "multi-agent software company" packaging — product managers, architects, engineers, and QA collaborating — is entirely absent from this Android testing demo. The entire module defines only one custom Role class ("role," meaning an AI unit capable of executing tasks autonomously). The so-called "role division" simply doesn't exist.
What this is
MetaGPT is one of China's best-known open-source multi-agent frameworks. What got analyzed this time is an unassuming sub-module in its repository — AndroidAssistant — which can autonomously explore or learn from human demonstrations to complete phone operation tasks.
Walking through the source code, you'll find the demo defines only one AndroidAssistant Role, relying on an if/elif branch to assemble different Actions ("actions," meaning specific sub-tasks called by a Role) under different configurations. No product manager, no engineer, no QA exchanging messages with each other.
What's actually reused is the Role/Action/Team scheduling skeleton — the per-round driver loop, control over whether memory carries into the next round, and whether Actions execute once or repeatedly. This infrastructure, originally designed for multi-agent collaboration, has been repurposed for a single testing Agent.
In plain language: the star product's "multi-employee collaboration" narrative didn't land. What landed is the underlying loop, memory, and orchestration code.
Industry view
Positive voice: This is the "down-to-earth side" of open-source frameworks. A general-purpose framework shouldn't be forced to deploy all of its capabilities in every scenario. Letting the scheduling skeleton be reused and letting a single demo run end-to-end is itself a demonstration of engineering strength — there's no need to force it to play "software company" for marketing's sake.
Opposing view / risks: What deserves more vigilance is the phenomenon of "packaging detached from substance" itself. When a company consistently sells products with narratives like "AI software company," "AI employees," and "multi-agent collaboration," yet its source code only exposes the scheduling skeleton — enterprise IT decision-makers paying for it can easily buy PowerPoint decks instead of real capabilities. The complexity of true multi-agent collaboration (message distribution, state conflicts, debugging difficulty) is far higher than a single Agent. If the claimed collaboration capabilities lack substance, enterprises pay collaboration prices for solo-execution capability.
Our judgment: In open-source projects like these, the louder the "role division" narrative, the higher the ROI of auditing its source code — it can directly tell you what the vendor itself actually believes in versus what users actually receive.
Impact on regular people
For enterprise IT: When evaluating AI automation products, push vendors to translate "multi-agent collaboration" into which specific module is sending messages to which other module. If they can't answer, they're probably selling a concept.
For individual careers: Within the next year, "AI employees" and "agent collaboration" will increasingly show up in the SaaS you purchase — knowing the difference between a "scheduling skeleton" and a "role narrative" can help you spot which are real tools and which are packaging.
For consumer markets: Has little to do with you for now; mobile test automation is still an engineer-circle affair. The conclusions here won't spill outward until AI testing capabilities actually land inside consumer-grade Apps.