I keep seeing the same architecture work across very different problems. It is not a mega-agent. It is not a chatbot taped to an API. It is a fabric of small, specialized programs — each one doing exactly one thing with high confidence — coordinating through shared context.
The mega-agent fantasy is seductive. One model, one prompt, one black box that takes your question and produces an answer. It works in a demo. It collapses in production the first time it has to do anything that matters because you cannot tell which step it got wrong, and you cannot trust the output enough to act on it.
The mesh is the opposite. An enrichment agent qualifies an inbound visitor. A routing agent decides where they go next. A content agent rebuilds the experience around them. An analytics agent watches all of them and feeds learnings back. Each one is small enough to inspect, narrow enough to be confident, and replaceable on its own. The same shape applies to operations — a data-collection agent, a prioritization agent, a delivery agent, a capture agent. Different domain, identical pattern.
What makes the mesh work is not the agents. It is the discipline around them. Context flows through written state. Each agent observes, applies its rule, hands off. Humans review the outputs that change the strategy, not every keystroke. The agents do the integration work. The humans do the judgment work.
I've argued elsewhere that the ERP is dead and that the gap between ad click and revenue is a mesh problem. The mesh is the architecture under both. Once you see it, you cannot unsee it — every enterprise problem worth automating has the same shape.
— Ry
Sources
Related Essays
The Mesh, Not the Monolith
One mega-agent that handles everything is exhilarating to demo and chaotic in production. Enterprise wants a mesh of specialized agents with human pilots.
The ERP Is Dead. The Agent Is Your Operating System Now
The ERP was supposed to solve the toggling-between-six-tools problem. It did not. The agent does — by becoming the operating layer on top of every system you already paid for.
The Atomic Agent Mesh: Architecture, Build-vs-Buy, and the Review Layer
Enterprise AI will not be one mega-agent. It will be a mesh of atomic, auditable units, and the companies that nail review and context will own the next infrastructure layer.
Key takeaways
- One mega-agent that tries to do everything fails in ways that are hard to debug and impossible to trust.
- The pattern that works is a mesh of small, specialized agents that each do one job with high confidence.
- Agents coordinate through shared context — observe state, apply logic, hand off, write back.
- Humans review the outputs that matter, not every step. That is what keeps the system running in production.
FAQ
Why not just use one big agent?
Because debugging a mega-agent is hopeless. When it makes a mistake you cannot tell which step went wrong, which prompt drifted, or which input confused it. A mesh gives you composable units you can inspect, replace, and version one at a time.
How do mesh agents coordinate?
Through shared context — written state, queues, or a working memory everyone reads from and writes to. Each agent observes the state, runs its specialty, and updates the state for the next agent. It is the same discipline that makes microservices work.