The agentic Fermi paradox: if agents are everywhere, where are they?

I keep hearing that agents are everywhere. I am still trying to work out where they are.

The thought takes me back to the Fermi paradox. Given the number of stars, planets and opportunities for intelligent life to emerge, Enrico Fermi reportedly asked a wonderfully simple question: where is everybody?

I have started to wonder whether agentic artificial intelligence has developed its own version of the same problem.

Walk around Cannes, Possible or almost any advertising conference and there is no shortage of agents. We have agentic platforms, autonomous workflows, intelligent assistants and a growing collection of acronyms to connect them all.

The conditions also look ideal. Advertising contains millions of repetitive decisions across planning, activation, optimisation, measurement and reporting. The industry has vast amounts of data, software-based workflows and constant pressure to reduce cost.

Building the basic technology is increasingly accessible. I spent a couple of evenings creating skills, connecting tools through Model Context Protocol and building agents in Codex and Claude. I am not pretending those experiments were production systems, but they worked. More importantly, they showed me how low the barrier to learning has become.

So, if agents are relatively easy to build and the potential applications are everywhere, why is the industry not visibly full of them?

The difference between assistance and agency

We have built plenty of useful assistants. We have built far fewer systems that companies trust to exercise genuine agency.

A chatbot that answers a campaign question may save time. A tool that recommends an audience may improve a decision. Neither is the same as a system authorised to select inventory, move budget or change an investment without somebody approving it.

This distinction gets lost rather easily because almost everything is now described as an agent.

I do not think the agentic universe is empty. I think many of its inhabitants are assistants wearing slightly more impressive name badges.

The agentic great filter

One explanation for the original Fermi paradox is the “great filter”: a stage that very few civilisations manage to pass.

Agentic systems have their own great filter between an impressive demonstration and dependable production use.

The model needs accurate data, access to the right systems, clear authority and defined limits. Somebody must decide what it can do, what it cannot do and when a person must intervene.

Then the less glamorous parts arrive. Legal wants to understand liability. Finance wants to know who authorised the spend. Technology wants to secure the connections. The client wants an explanation when the outcome changes.

Protocols can help systems communicate. They cannot decide whether a chief financial officer, client or regulator will trust the result.

The technical barrier to creating an agent is falling quickly. The organisational barrier to delegating authority is not, and the gap between the two may be the real agentic great filter.

The useful agents may be quiet

There is another possible explanation. Perhaps the most effective agents are not being announced.

A company that has found a genuine advantage in planning, pricing or optimisation has little reason to explain exactly how it works. The agent may simply sit inside an existing product or workflow without a launch event, a new interface or a person-shaped avatar.

We may therefore be looking for the wrong evidence. The better questions are fairly ordinary. Do campaigns launch faster? Can each person manage more activity? Do decisions improve? Does the cost of execution fall? Does work that was previously uneconomic now become worthwhile?

If an agent works properly, the customer may only notice that the service improved; the signal will be economic rather than theatrical.

Some agents may not deserve to exist

There is also a less exciting answer, although I suspect it is true more often than the industry would like to admit.

Some proposed agentic applications simply do not create enough value. Some decisions occur too infrequently to justify the integration. Some need judgement but not autonomy. Others can be handled more reliably through ordinary software and a decent set of rules.

Calling a workflow agentic does not improve its economics. The current enthusiasm has produced plenty of solutions searching for decisions they should be allowed to make. Are the demonstrations clever? Often, yes. Does that mean somebody will trust them with money or pay for the result? Not necessarily.

Where is everybody?

I think the answer is that most agents remain primitive, constrained or hidden. Some are useful assistants presented as autonomous systems. Some are experiments waiting for trusted data and permissions. Some probably operate quietly because their owners have no interest in advertising the advantage.

The limitation is not only model intelligence or whether systems can connect. It is whether an organisation is willing to let software exercise meaningful authority.

That is where the conversation becomes real. What decisions will the agent be allowed to make? Under whose rules? Who can stop it? Who takes responsibility when it gets the decision wrong?

Until companies can answer those questions, the agentic universe will continue to look strangely quiet.