CODE IN TOKYO

Do We Still Need the Multi-Agent Designs That Were So Painful One Year Ago?

The Multi-Agent Systems I Tried a Year Ago

About a year ago, I experimented with multi-agent systems in various ways.

At the time, people were saying that the next era would be one where AI agents cooperate with each other to get work done. Naturally, I tried building that kind of system too.

One agent would do research. Another would write. A final agent would review the result. Each agent had a role, and the output from one agent would be passed to the next.

As a concept, it was very easy to understand.

But actually building it was troublesome.

I had to write prompts for each agent, decide which information should be handed over, and control the execution order. If I ran agents in parallel, I then had to decide where and how to merge their results.

And despite all that effort, the system was not especially smart.

The context would drift halfway through. Multiple agents would start doing similar work. Sometimes they would repeat the same thing endlessly.

In the end, I started thinking, “Wouldn’t it be faster if I just gave instructions to AI step by step myself?”

So I moved away from multi-agent systems for a while.

Recently, however, I have started using agent features again.

And I had one thought.

Much of what humans were desperately designing one year ago is no longer necessary.

When You Ask for Research, Another Agent May Investigate in the Background

Search is one of the clearest examples in recent AI agents.

When I ask for a somewhat broad investigation, the main agent does not always try to handle everything by itself. Sometimes it delegates research to a sub-agent.

I did not give detailed instructions like “create a researcher agent.”

I simply explained what I wanted investigated.

In search-related workflows, multiple agents can investigate from different angles and then combine the results.

One year ago, the human side would have defined a role called “search agent,” written its prompt, and designed the mechanism for passing search results to the next process.

Now, humans do not need to design that much detail manually.

What was all that effort for?

Dynamic Workflows Can Handle More of the Parallelization

Parallel processing is similar.

Previously, humans had to think through which tasks were independent and build workflows such as “A and B can run in parallel, while C should run after A is complete.”

With recent Dynamic Workflow approaches, Claude can think about how to use sub-agents and create workflows that progress through multiple tasks in parallel.

Of course, this is not a complete black box. There is still a mechanism for humans to review the generated workflow.

Even so, the need for humans to design the execution graph in detail from the beginning has become much weaker.

I once opened three terminal windows and ran multiple AI sessions in parallel.

The result was that my human brain could not keep up.

Just checking which AI was doing what became tiring.

By parallelizing AI, I made the human the bottleneck.

Now, even that traffic control is gradually shifting toward the AI side.

We No Longer Need to Define “You Are the Reviewer” So Strictly

Roles also no longer need to be defined as strictly as before.

Of course, I am not saying that AI will create a perfect organizational chart without any instruction.

But if you explain that you need someone to review the work from a perspective separate from implementation, AI can now behave much closer to that role.

You do not necessarily need a long prompt such as:

“You are a senior software architect with 20 years of experience. From the perspectives of separation of responsibility, maintainability, extensibility…”

In many cases, “please review this” is enough.

What matters is not the job title, but what you want done.

When I think about it, human work may be similar.

Being told “you are a senior quality assurance specialist” is less clear than being told “look for the parts of this implementation that are likely to break.”

AI has started to understand context at that level.

What Was Last Year’s Multi-Agent Design Effort?

This can feel a little empty.

One year ago, I was thinking hard about agent roles, adjusting prompts, and building workflows.

Now, “research this,” “move ahead in parallel where possible,” and “review it from another perspective at the end” can already get quite far.

I do not think the work from that time was useless.

Because I had that experience, I can now imagine what AI is doing behind the scenes.

But if you asked whether I would build the same thing myself today, I probably would not.

When you build a complex structure around AI, that structure may become outdated before the AI itself does.

I have seen this pattern too many times recently.

Something humans spent six months carefully designing becomes a normal capability in the next model or a standard feature.

In the AI era, technical debt is not only dirty code.

It may also be technical debt when humans painstakingly build something that AI will soon handle by default.

If that is true, future system design should probably stay as simple as possible.

Humans should only fill in the missing parts, and when AI becomes smart enough to make some structure unnecessary, we should throw that structure away.

One year ago, I struggled with multi-agent design.

Today, I ask AI, “Split the work if needed.”

And unfortunately, today’s approach works better.

Contact

I provide system development and technical consulting using AI, AWS, and Claude Code.

Contact form: https://holly-money-e94.notion.site/390ff30cf18c8086a676fe630d171873