What would Darwin say about AI Agents?

The cooler kids in the AI world are going wild for Manus (or at least they were last month). For those who aren’t familiar, Manus makes bold claims — namely that it is able to do everything (from app development to teaching and a lot in between). They aren’t the only company making claims, but they seem to be the best (based on the hype).

Maybe we will have a Single Agent to Rule them All.

But everything from evolutionary biology would indicate otherwise…

We don’t have super powered animals or humans that are the best at everything.

What can nature teach us about agents?

The Path to Agentic Stasis

Animals change constantly. Helpful changes are preserved, bad ones are not (through reproduction and death).

Horseshoe crabs have been doing their thing the exact same way for almost 450 million years. Genetic mutations don’t make them more successful in the costal waters where they live, so any changes aren’t preserved. Over time, the horseshoe crab’s “features” have been extraordinarily stable.

Agents will be the same. We will get some agents that are really good at performing certain tasks (like scheduling between two calendars). They will probably stabilize pretty quickly. More complex agents (like therapists) will stay dynamic for a while. There are a lot of capabilities they will need to “evolve” to solve problems.

In evolutionary biology, when an animal fits so well into a niche that it stops dramatically changing, this is called reaching stasis. Some agents will reach stasis faster, some may never get there.

MCP & A2A started an Agentic Explosion

What causes rapid evolutionary change? Pressure! And it just got here in the form of protocols…

Two frameworks are getting a lot of chatter lately: Model Context Protocol (MCP) and Agent 2 Agent (A2A). While one is designed for tool calling, and the other for inter-agent communication, the gap between what is a “tool” and what is an “agent” won’t exist for that much longer.

Both turn what’s happening on the other side of the protocol into a black box. You can say things like “book me a dinner with Joe” and the agent (or tool) will do exactly that.

These black boxes are going to lead to an evolutionary explosion because it makes it so much easier to coordinate a fleet of agents and tools. You tell an agent or tool to book the dinner, and it figures it out. You know nothing about its internals while it is operating.

This is going to lead to a lot of specialization, and quickly.

What is an agent anyway?

That which we call an agent by any other name would work as sweet

Tools are closer to horseshoe crabs — they are really good at one thing.

Agents are more like Darwin’s finches — they are adapting themselves constantly to solve a particular task, trying different tools to get to their end state.

Over time, agents will harden. People don’t care how the agent accomplishes its task — they just want it to happen as quickly (and low cost) as possible. Once agents discover the fastest path (and execute it thousands to millions of times) — they will turn more “dynamic” tool calling into deterministic code. It is cheaper and faster to be more predictable if you are solving the same problem over and over.

Why does this matter? Evolution isn’t constant and gradual. There are moments of relative stability and moments of extreme divergence.

This is theory of Punctuated Equilibrium.

We are in one of those extreme divergence moments right now in AI. When these moments happen, you want to be evolving.

Your company should be evolving, your product should be evolving and YOU should be evolving.

Work and life today is not the same as what it will be in the near future. So get on board or you will be left behind.

If there was ever time to fill a niche — it is today.

The good news is all of this “new” stuff is really brand new. The world’s expert at using these new, hot-shot protocols has been using them for less than a single semester. Go for it!

If you can find a way to harden an agent and deliver better than others for a particular business problem… you will probably become the tool for solving that problem. There’s a good chance if you do it well enough, you’ll be that tool provider forever. As long as you don’t get greedy. It’s a great time to be that first horseshoe crab.

Looking into the future… what is a tool anyway?

What will we call an agent that has hardened into a crab? When it’s no longer as dynamic. When it just gets the right answer… and it does that every single time. And quickly.

At that point, it still an agent? Or has it become a tool?

Does an agent need to have an LLM acting erratically to still be an agent? Does it need to be slow?

If it does not, do we even need the Agent 2 Agent protocol, or can we just start treating agents today as tools with semantic text input?

This one thought that has kept me up the last week:

What if we just had all agents respond via tool calls in MCP and skip A2A? That feels like the inevitable end state...

After all, MCP Servers are long lived sessions with state. And I don’t know what’s behind the black box.

Do I care if what I am using is an agent or a tool?

Not if it just solves my problem.

Just do the thing I want. End of story. Give me a horseshoe crab.

Side Note: Am I the only one?

When using agents, I feel like longer running agents are better.

That’s just my expectation. Deep Research is good because it takes a long time. That’s it thinking and that’s why the answer is better. It has to be. If it came back with an answer right away, it must have not done its job correctly.

But what if I’m wrong? What if it had the answer all along and all of Deep Research a UI trick to make us feel like these answers are better? I often get the same “answer” from shorter running agents and a few queries.

That trick works on me today. But in the future, maybe my expectations will change. Time will tell.

I certainly don’t think Cursor is going to be better when it makes me wait.