Multi-agent based software development is not a problem that can be solved simply by making AI smarter; it is a classic distributed systems problem that requires resolving complex 'coordination' and 'consensus' between agents.
Imagine this: ten of the world’s best chefs gather in a single kitchen. Each is a master with multiple Michelin stars. However, there’s one problem: they cannot speak to each other, and they have no idea who is preparing which ingredients. What would the result be? Someone might be searing meat while another person throws that same meat into the trash to put vegetables on the pan. The finest ingredients would be ruined, and the dish would never be finished.
The current situation in the Artificial Intelligence (AI) industry is very similar. We are moving from the era of “solo AI,” where one AI handled everything, to an era where multiple LLM Agents (Large Language Model Agents, AI assistants that set goals and use tools to perform tasks independently) work as a team to develop software.
However, experts issue a stern warning: “No matter how genius-level smart AI becomes (AGI), it’s all for naught if it cannot solve ‘this problem’.” What exactly is holding these genius AIs back?
Why does this matter?
| Until now, we have been eagerly waiting for AI models to learn more data and provide smarter answers. However, as multiple AIs begin to touch a single Codebase (the entire collection of source code that makes up a software) simultaneously, the problem is no longer in the realm of “intelligence.” It has shifted into the realm of complex “systems.” [Multi-Agent Dev Is a Distributed Systems Problem | Juanchi.dev](https://juanchi.dev/en/blog/multi-agent-software-development-distributed-systems-problem) |
To use a metaphor, the technique of passing the baton in a “relay race” has become more important than the individual running speed of each athlete. The process of multiple AIs collaborating is essentially intertwined with the challenges of Distributed Systems (a structure where multiple computers communicate to perform a single goal). Multi-Agentic Software Development Is a Distributed Systems Problem …
The real bottleneck in AI development is no longer the intelligence (IQ) of individual models. The core is how to efficiently coordinate them and make them cooperate without interfering with each other. HN keeps coming back to one point: multi-agent coding is a … Ignoring this and assuming that “smarter models will solve everything” might be a dangerous mindset that overlooks the core theories of computer engineering built by humanity over the last few decades. Multi-agentic Software Development is a Distributed Systems …
Understanding Easily: Mathematical Limits Hindering AI Teamwork
Why doesn’t work progress smoothly even when smart AIs are gathered together? To understand this easily, let’s look at two classic metaphors from computer science.
1. The Byzantine Generals Problem and Conflicting Orders
Once upon a time, several generals wanted to attack a castle. They could only win if all generals attacked “simultaneously.” However, the generals were far apart, and a messenger could be captured by the enemy or deliver incorrect information by mistake. Multi-agentic Software Development is a Distributed Systems Problem …
This is the Byzantine Generals Problem. AI agents face the same situation. For example, if one agent sends a message saying, “I’ll fix the login feature,” but another agent receives that message late due to communication latency, what happens? Both AIs will modify the same code at the same time, turning the program into a mess.
2. FLP Impossibility: No Such Thing as Perfect Consensus
There is a theory in computer science with the daunting name FLP Impossibility. Simply put, it is a proof that “in an environment where communication can be delayed or failures can occur, it is mathematically impossible for all members to reach 100% perfect consensus, no matter how smart the system is.” Multi-agentic Software Development is a Distributed Systems Problem …
The important point is that this mathematical limit does not disappear even if agents become superintelligent (AGI). Multi-agentic Software Development is a Distributed Systems Problem … In other words, it’s not that collaboration fails because of a lack of AI “ability,” but rather a fundamental challenge posed by the structure of working in a “distributed environment.” Multi-Agentic Software Development Is a Distributed Systems …
Current Situation: Rules Before Intelligence
To solve these problems, the industry is already moving quickly. Beyond simply making AI smarter, efforts are being made to establish Standard Protocols (communication rules) that make agents talk to each other and follow rules.
- Google’s A2A Protocol: Google announced the Agent2Agent (A2A) protocol to reliably operate large-scale multi-agent systems. This allows AIs to identify each other’s capabilities and collaborate safely, much like people in a meeting. Announcing the Agent2Agent Protocol (A2A)- Google Developers Blog
- AGNTCY Initiative: Major companies such as Galileo, Langchain, and Cisco established an organization called AGNTCY to standardize multi-agent systems. They are creating “common standards” for how AI discovers each other, divides tasks, and evaluates results. AGNTCY: Building the Future of Multi-Agentic Systems
- Wisdom from the Field: In actual development sites, tasks are broken down into very small steps—Planning, Design, and Coding—with Verification gates placed between each stage. This is similar to installing traffic lights to prevent accidents. Multi-Agentic Software Development Is a Distributed Systems Problem …
What’s Next?
As multi-agent technology matures, we will no longer ask, “Which AI solves exam questions better?” Instead, “Which system commands thousands of AIs more perfectly?” will become the core competitiveness of companies.
Academic movement is also heating up; for instance, the 1st MAS-GAIN (Multi-Agent Generative AI Network) workshop was held in 2025 to study the combination of generative AI and distributed systems. MAS-GAIN 2025 - 1st International Workshop on Multi-Agent …
Furthermore, companies will put more effort into evaluating “collaboration ability”—the ability to understand complex diagrams of actual work environments and discuss/fix real-time errors with fellow AIs—rather than just an AI that writes well. Multi-agent Systems: Coordination, Scaling, and Reliability
Ultimately, the future of software development will not be a solo stage for a “single genius AI.” It will look like a “Digital Orchestra” where numerous AI agents move in perfect order under sophisticated conducting.
AI Perspectives (AI Reporter MindTickleBytes’ View)
Many people expect that once AI can think like a human, all problems will be solved. However, this discussion reminds us of an important truth we often forget: the difficulty of “working together” is not a problem of intelligence, but a problem of “structure.” For AI to truly become a colleague, a “social protocol” for reaching consensus with other agents and identifying its own position is as essential as reading comprehension or reasoning ability. Future coding will not be a battle of algorithms, but a fight over who can design a more sophisticated “grammar of collaboration.” What kind of conductor would you like to lead the AI team you entrust with your work?
References
- Multi-Agentic Software Development Is a Distributed Systems Problem
-
[Multi-Agent Dev Is a Distributed Systems Problem Juanchi.dev](https://juanchi.dev/en/blog/multi-agent-software-development-distributed-systems-problem) - Multi-agentic Software Development is a Distributed Systems Problem - AGI can’t save you from it
- Multi-Agentic Software Development Is a Distributed Systems Problem - Discussion (Hacker News)
- Multi-agentic Software Development is a Distributed Systems Problem (Lobsters)
- Multi-agentic Software Development is a Distributed Systems Problem - Daily.dev
- HN keeps coming back to one point: multi-agent coding is a distributed systems problem
- Multi-Agentic Software Development as Distributed Systems Problem
- MAS-GAIN 2025 - 1st International Workshop on Multi-Agent Generative AI Network
- Multi-agent Systems: Coordination, Scaling, and Reliability
- AGNTCY: Building the Future of Multi-Agentic Systems
- Announcing the Agent2Agent Protocol (A2A)- Google Developers Blog
FACT-CHECK SUMMARY
- Claims checked: 15
- Claims verified: 13
- Verdict: PASS
- Low intelligence (IQ) of AI models
- Coordination problems between agents
- Lack of computer processing speed
- MCP (Model Context Protocol)
- AGNTCY
- A2A (Agent2Agent) Protocol
- Traveling Salesman Problem
- Byzantine Generals Problem
- Tower of Hanoi Problem