The AI agent moment has not (yet) arrived

It's Thursday, 13^th August 2026. Hello and welcome back to Bold Efforts!

There is something strange about the current state of AI.

ChatGPT, Claude, and Gemini have clearly crossed into everyday life. People use them to understand unfamiliar topics, improve their writing, compare choices, prepare for meetings, learn new skills, and think through decisions. AI chat is no longer limited to software engineers or people who closely follow technology.

AI agents have not made the same leap somehow.

This is surprising because the underlying technology appears to be arriving quickly. Models can browse the internet, operate software, write and execute code, call external tools, and complete increasingly complex tasks with limited supervision. Every large technology company is building some form of agent platform. Startups are building agents for sales, finance, travel, hiring, customer support, legal work, healthcare, and almost every other industry.

And yet, outside technology circles, very few people and companies seem to have an agent they regularly rely on.

Most people I know use AI to help them think. They do not often allow it to act.

They might ask ChatGPT to compare two insurance policies, but they will still choose the policy themselves. They may use Claude to prepare questions for an important meeting, but they will not ask it to conduct the conversation. They may use Gemini to plan a trip, but they will still open multiple tabs, check the hotels, compare the flights, and make the bookings.

AI has become useful without becoming responsible. That distinction may explain why the agent moment has not yet arrived.

The most common explanation is that agents are still unreliable. They make mistakes, lose context, misunderstand instructions, and sometimes take actions that a person would never approve.

That is true. But it is probably not the full explanation.

The problem is not only whether an AI system can complete a task. The problem is whether a person is comfortable transferring responsibility for that task.

There is a meaningful difference between asking for a recommendation and allowing software to make a decision on your behalf. A recommendation can be reviewed, ignored, or modified. An action creates consequences.

Imagine asking an AI assistant to organize a meeting with five people. The task sounds simple. In practice, it needs to know which meetings are important enough to protect, whose calendar should take priority, whether an early morning slot is acceptable, how much travel time to leave, and whether a particular person should be included at all.

A human assistant often learns these preferences gradually. Some rules are explained directly, but much of the judgement comes from observation, correction, and experience.

An AI agent usually starts with none of that context.

The same problem appears across ordinary work. Asking an AI to draft a response to a client is straightforward. Asking it to send the response is different. It now needs to understand the history of the relationship, the sensitivity of the situation, what can be promised, what should remain unsaid, and when the issue should be escalated instead.

The difficult part is the judgement surrounding the action.

Delegation is often described as a way to save time, but delegating well requires work.

You need to explain what you want, provide the relevant context, define the boundaries, review the result, and correct mistakes. Managers experience this constantly. A task may take twenty minutes to complete but thirty minutes to explain properly, so they continue doing it themselves.

AI agents face the same problem.

If a person has to carefully describe every preference, connect several accounts, grant permissions, monitor progress, and review each action, the agent may not feel like automation. It feels like another system that needs to be managed.

This is one reason AI chat has spread more quickly. The interaction is simple and familiar. You ask for help, receive a response, and remain in control. The software does not need deep access to your life or work. A poor answer is usually easy to recognize and cheap to ignore.

Agents require a much larger leap of trust.

This leap becomes harder because many everyday tasks are not clearly defined in the first place. People often discover what they want while looking at the available options. Preferences change depending on context. The correct decision may depend on information that has never been written down.

The user is not always withholding a precise instruction. Sometimes no precise instruction exists.

The growth of coding agents can make the broader agent shift appear closer than it is.

Software development is one of the few environments that is already well suited to delegation. Most of the relevant information is available inside the computer. The task can often be stated clearly. The agent can inspect the codebase, make a change, run tests, read logs, and evaluate whether the result worked.

Mistakes are also relatively easy to contain. Changes are recorded. Previous versions can be restored. The agent can try something, observe the result, and try again. The environment provides feedback.

Most office and consumer tasks do not work this way. There is no automatic test that determines whether a presentation makes the right political argument. There is no clean measure of whether a supplier negotiation was handled well. There is no simple rollback after an agent shares confidential information with the wrong person.

This does not mean agents will remain limited to software development. It means coding agents benefit from conditions that do not yet exist in most other areas.

The code is structured. The tools are connected. The desired outcome is often visible. The consequences of mistakes can usually be reversed. The real world is far less organized.

Companies often imagine that agents can be placed on top of existing processes and immediately make them faster.

But many business processes are not really systems. They are collections of habits, spreadsheets, inboxes, informal approvals, and knowledge held by particular employees.

A process may appear simple from the outside because an experienced person knows how to handle the exceptions. They know which customer requires a different approach, which number in the spreadsheet cannot be trusted, which manager needs to approve an unusual request, and which written rule is rarely followed in practice.

None of this may exist in the official documentation.

An agent operating inside such a company does not only need access to information. It needs to understand which information is current, which source is reliable, which rules are flexible, and which exceptions matter.

This is why many agent pilots look impressive during a controlled demonstration but struggle during normal use. The demonstration contains a clean task, a prepared environment, and a limited range of outcomes. Real work introduces missing information, conflicting instructions, changing priorities, and unusual cases.

The model may be capable. The environment around it is not ready.

There is another reason the agent moment may be taking longer than expected.

“AI agent” is a technical category, not a customer need.

People do not wake up hoping to adopt an agent. They want to avoid spending two hours resolving a billing problem. They want to find a suitable job without reviewing hundreds of irrelevant listings. They want their monthly accounts prepared correctly. They want a delayed shipment fixed without making six phone calls.

The successful product will be the one that solves the problem. Whether an agent sits behind it may not matter to the user.

This is how many important technologies spread. Consumers did not demand cloud computing, recommendation systems, or computer vision. They adopted products that happened to use those technologies to deliver a better experience.

Agents may follow the same path. The strongest agent products will probably be narrow before they become broad. They will operate inside a clearly understood domain, use reliable information, and begin with actions where the downside is limited.

At first, the system may recommend. Then it may prepare the next step. Later, it may act with approval. Full autonomy will come only after the product has repeatedly demonstrated that it understands the user and the environment.

Trust will be built gradually, not requested all at once.

Many agent products still assume that the user wants to manage work through a conversation.

Chat is useful when the person is still figuring out what they want. It is less useful when the task is repetitive, structured, or visual.

A person comparing financial products may want to see a table. Someone planning a week may prefer a calendar. A manager reviewing candidates may want a shortlist with clear reasons, not a long conversation with an assistant.

The best agent products may use conversation only when conversation is the most natural way to express intent. The rest of the experience may look like familiar software, with the agent quietly handling work behind the interface.

This matters because a new technology does not automatically require a new behaviour.

People adopted AI chat quickly because asking questions was already familiar. Agents will need to fit into existing habits just as naturally. Requiring users to move every task into a conversational interface may create friction rather than remove it.

It is possible that AI agents will never have a single, visible breakthrough moment.

There may be no one product that suddenly makes everyone talk about agents. Instead, they may spread quietly through software that solves specific problems well.

People may notice that their expenses are processed automatically, customer requests are resolved faster, meetings require less preparation, or a complex application takes minutes instead of hours.

They may not describe any of this as using an AI agent.

That would not make the shift less significant. It would simply mean that agents succeeded by disappearing into useful products rather than becoming a consumer category of their own.

The technology is increasingly capable of acting. But widespread adoption requires more than capability. Agents need reliable information, clear boundaries, useful interfaces, reversible actions, and enough context to exercise judgement.

Most importantly, they need to solve something people already care about.

The AI agent moment has not yet arrived. When it does, it may not begin with people deciding they want agents.

It may begin when they stop noticing that an agent is involved.

Until next week,
Kartik

I write Bold Efforts every week to think clearly about where work and life are actually headed. If you want these essays in your inbox, you can subscribe here.