AI Agent
A single AI processes in sequence
Processes run in sequence. The next process does not start until the previous one has finished. The structure is simple and the progress of processing is easy to trace, but it takes time.
Architecture
Even when they are all called "AI," there are three ways of building it. What each can do, how fast it processes, and the impact when a fault occurs all differ. We use them selectively according to the task.
Comparison
The diagram shows how the same job, made up of multiple steps, progresses when processed in three different ways of building it. The horizontal axis is the passage of time. A lit frame indicates a step in progress, and a green mark indicates that the job is complete.
AI Agent
Processes run in sequence. The next process does not start until the previous one has finished. The structure is simple and the progress of processing is easy to trace, but it takes time.
AI Orchestration
A central mechanism manages the order of processing and the re-execution of failed processes. Because processes that call external services or run at set times are handled in the same flow, the point where a process stopped can be identified immediately.
AI Multi-agent
Multiple AIs with different roles process work at the same time, and a central AI cross-checks the results and finalizes them. Processing becomes faster, but designing the part that combines the results becomes more difficult.
This diagram shows the difference between the two approaches. It does not measure actual time required.
Three ways
It is not chosen on processing speed alone. It is chosen by considering the nature of the work and what impact a fault would have.
It takes on a defined role, uses tools and data as appropriate, and advances the procedure on its own. It is a way of building that entrusts one person's work to the AI through to the end. Because what it did along the way is kept in a single record, it is also easy to trace causes.
AI, external services, and processes that run at set times are managed centrally and executed in a defined order. The central system also handles re-running failed processes and sending notifications when something stops. We use this approach for work that runs while no one is watching.
Multiple AIs with different roles share the work, and a central AI compiles the results and finalizes them. A person makes the final decision. Work that deals with large volumes of material open to differing interpretations cannot be completed unless it is divided up.
How to choose
If the procedure is fixed and the volume is what one person can handle, an AI agent is sufficient. Because the structure is simple, it is also quick to fix. There is no need to build something large from the start.
If external services, scheduled processing, or handoffs to other systems are involved, build it with centralized management. Being able to identify immediately where processing has stopped is useful in day-to-day operation.
When handling materials that can be read in different ways and the volume is large, we divide roles and proceed in parallel. On the side that consolidates the results, we assign a role of finding discrepancies between the results of the AIs. A person makes the final decision.
There is no need to commit to a single way of building. We start small and move only the parts whose volume has grown to a division of labor among multiple AIs. We design the system so that it can be adapted by replacing parts, without rebuilding the whole.
In production
The explanations here are based on systems we actually operate.
8systems
AI systems we actually run in our own business
7agents
The agents that make up the multi-agent system overseeing the operation of the entire business
86functions
In-house system functions that can be run directly from AI assistants
Values observed from primary sources (as of November 6, 2026)
How this site is operated