The next step with AI is agents that work without you

IndyDevDan·October 26, 2026·Watch the video

AI agents have become so good at programming and problem solving that humans can increasingly become the bottleneck. Even so, many people still use AI the same way: they write an instruction, wait for the result, check it and give the next instruction. Even with powerful models that can solve complicated tasks, it is still the human who has to keep the process going.

The message of the video ELITE Agentic Engineers Are LEAVING THE LOOP is that the next step in AI development is not mainly about writing better instructions or using even more powerful models. It is about building systems that let AI agents work independently, even when you are not there.

Although the video is mainly about programming and software development, the principles are far more general. The same way of using AI agents can be carried over to purchasing, logistics, customer service, finance, marketing and many other tasks.

This requires a different way of thinking. Instead of using AI to carry out one task at a time, you build systems that can start tasks, carry them out, check the results and keep working without constant human follow-up.

From managing every task to monitoring the results

The development is described through three levels of human involvement. 1:37

In the loop: You are an active part of the whole process. You start the task, write instructions, keep an eye on the agent's work and decide what happens next. This is still a common way to use AI coding agents such as Codex and Claude Code.

Near the loop: The AI agent does most of the work itself, but involves you in important decisions. For example, it can detect an error, investigate the cause, suggest a fix and ask for approval before the change is put to use. You don't need to follow the process, only make a decision when it is actually needed.

Out of the loop: The agent carries out tasks automatically within predefined limits. It can start, carry out and check the work without you having to approve each step. Your role is more about keeping an eye on the results and stepping in when needed.

The goal is to move more and more tasks from the first level to the last. This does not mean that human control should be removed from every process, but that control should be used where it adds value.

An important point in the video is that the transition does not have to happen all at once. Near the loop is a natural step in between. The agent can work independently on most things, while the human still keeps control over critical decisions.

Four basic skills become essential

To build systems like these, you need to master four key areas. 15:39 The video calls them The Core Four, and they form the foundation of how an AI agent works.

Context

Context is about what information the AI agent has available when it is going to carry out a task. 17:05

An agent that is going to fix an error on a website, for example, needs to understand how the project is built, which files are relevant, which rules apply and what has been done before.

In the same way, an agent that handles customer orders needs access to relevant order information, supplier agreements, inventory status and delivery terms.

Giving the agent as much information as possible is not necessarily best. What matters is that it gets the right information at the right time. This is called context engineering.

The better the context is organized, the better the chance the agent has of making the right decisions without having to ask the human along the way.

Model

Choosing a model is about using the right AI model for the right task. 17:12

Some models are especially good at complicated programming and problem solving, while others are faster and cheaper. A simple task such as classifying an error message or checking a delivery status does not necessarily need the same model as one that has to plan and carry out large changes to software.

This is also where costs become important. The video calls this token economics, meaning the relationship between how much the AI model costs to use and what value it delivers. AI models split text and other information into tokens, small units that can be whole words, parts of words or parts of sentences. The word "unbelievable" can, for example, become several tokens.

When agents work continuously and carry out many tasks, unnecessary use of expensive models can get costly. Good AI development is therefore also about finding the right balance between quality, speed and price.

Instructions (prompt)

Instructions are about how the agent is told what to do, which rules to follow and how to handle different situations. 17:20

Here it is important to tell the difference between an ordinary user message and a system instruction. The user message usually describes the specific task, while the system instruction defines the agent's role, limits and basic behavior.

An agent that is going to work independently needs clear instructions on how to check its own work, when to stop and when to ask for help.

It is not enough just to ask an agent to "follow up on all delayed deliveries." It also needs to know how delays should be identified, which actions it is allowed to take, when the customer should be contacted and when an employee has to be involved.

Tools

An AI model can reason and suggest solutions, but it needs tools to take actions in its surroundings. 17:35

These can be access to files, a terminal, a browser, email, databases, finance systems, order systems or other AI agents. An API is an interface that lets programs communicate and exchange information automatically.

Building and adapting the system that gives the agent these abilities is called harness engineering. Among other things, it covers which tools the agent gets access to, how tasks are carried out and which safety mechanisms limit what it can do.

This is one of the video's most important points. The difference between an ordinary AI agent and a specialized, independent agent lies not only in the model, but in the whole system around it.

AI agents that work before you ask them to

A central part of this development is what are called durable agents, or persistent AI agents. 6:07

These are agents that can work over long periods, keep information about their tasks and continue working after interruptions. They can remember earlier decisions, handle errors and resume processes after restarts.

Agents like these often run in their own isolated environments, often called sandboxes. This means they get access to computing power, files and specific tools in a limited environment. With the right security rules, they can work without getting unlimited access to the rest of the system.

If the environment runs on a server or in the cloud, the agent can also keep working even if you close your computer.

But the most important thing is how the agents start tasks.

Instead of the human always having to write an instruction, an agent can be activated by a specific event. That could be a new error message, an update in a database, an incoming customer order or a set time of day. A scheduled automatic start is often called a cron job.

The difference is that the agent no longer waits for you to start the work. It can start when the need arises.

Three concrete examples of independent AI agents

The video shows several examples of how systems like these can already be built. 8:43

The first is an agent that monitors software for errors. When something goes wrong, it can examine the logs, find a possible cause, write a fix and test the solution. Then it presents the proposal to a developer who can approve the change. The work before and after the approval can be automated.

The second example is a news agent that continuously collects information, filters out irrelevant content and assesses which stories deserve attention. Users do not have to search through large amounts of news themselves. The agent does the groundwork and presents information that may be worth looking into further.

The third example is about building user interfaces from a reference image. The agent gets an image of how a web page should look, writes code, takes a screenshot of the result and compares it with the original. Then it makes improvements and repeats the process.

This is an example of an automated improvement loop, where the AI agent uses feedback from its own tests to get closer to a defined goal. The video demonstrates a system like this, but the results also show that the agent does not always manage to recreate the design perfectly.

What they have in common is that the agents do much of the work on their own and involve the human when needed.

The same principles can be used in almost every part of a business

Although the examples in the video come from software development, there is no reason to limit this way of thinking to programming. Many tasks in a business already follow defined processes that can be fully or partly automated with AI agents.

Take purchasing and logistics as an example.

A company that sells IT equipment may have hundreds of orders and deliveries to follow up. Some items are shipped from its own warehouse, while others are shipped directly from the supplier to the end customer.

Traditionally, this takes a lot of manual work. An employee has to check order confirmations, check delivery dates, follow up with suppliers, get tracking information and inform customers if something is delayed.

With the right access to the company's systems, an AI agent can take over large parts of this process.

When a new order comes in, the agent can check the details, look up inventory status and prepare an order with the right supplier. Once the order is confirmed, it can follow the delivery, get the tracking number, check the expected delivery date and keep the customer updated.

If the supplier does not ship the item as agreed, the agent can detect the deviation, contact the supplier, look into alternative solutions and alert the responsible employee when a decision is needed.

Simple, clearly defined tasks can be carried out automatically, while larger purchases, price changes and unusual delivery agreements, for example, may require approval.

This is a made-up example of how the technology can be used, not a claim that all of today's AI agents already handle processes like these without errors. The company still has to make sure it has correct data, the necessary integrations, clear authorizations and control mechanisms.

The same principle can be carried over to customer service, invoice checking, finance, HR follow-up, marketing and many other areas.

The key question is not which industry you work in, but which tasks can be carried out without a human having to direct every single step.

Ready-made products such as Grok Bot and dots in ChatGPT

This development is also not only about systems that programmers have to build themselves. We are already seeing examples of AI products that make it possible to delegate tasks to persistent agents through ordinary conversation.

Two current examples are Grok Bot and dots in ChatGPT.

Grok Bot is built around the idea of AI agents as digital coworkers. The agents have access to their own work environments in the cloud and can use programs, websites and connected services. They can keep context over time, work with other agents and carry out tasks without the user having to follow every step.

OpenAI describes dots in ChatGPT in a similar way. A dot can be given responsibility for a project or a specific area, use connected tools, follow up on tasks over time and come back with results or questions that need human judgment.

Both products show how AI agents are moving from being tools you give one-off instructions to, toward becoming systems you can delegate ongoing responsibility to. They have different features and limitations, but they build on much of the same thinking.

What is also interesting is how the interaction changes. Instead of managing many different files, instructions and automations by hand, you can to a greater extent describe an area of responsibility in natural language, give the agent the access it needs and then follow up through conversations.

It is more like delegating tasks to a coworker than operating a traditional computer program. You describe the goal, explain the expectations, give feedback and let the agent carry out the work within agreed limits.

For more specialized tasks, such as managing purchasing processes or coordinating deliveries between several suppliers, it may still be necessary to build your own integrations and control systems. But the way of working itself is the same.

From doing tasks with AI to building systems that do them

This development also changes what it means to be a skilled developer, manager or professional.

As more and more people can use AI to write code, analyze information and carry out tasks, it becomes more important to understand how to put together complete systems. The value shifts from managing each individual task to developing processes, control mechanisms and structures that make the agents more independent.

The video draws a clear line between vibe coding and agentic engineering. Vibe coding is described as letting AI build software without necessarily understanding everything that happens along the way. Agentic engineering, on the other hand, is about understanding the system well enough that you can trust how it behaves within defined limits.

The same principle applies outside programming. A purchasing manager does not need to be able to write code to understand how a purchasing process works. On the contrary, expert knowledge about suppliers, prices, delivery times, risk and customer needs is essential for defining how an AI agent should work.

So professional expertise does not necessarily disappear when AI takes over more tasks. It gets a different role. People use their knowledge to define goals, rules, quality requirements and exceptions, while the agents can carry out more and more of the operational work.

A central principle from the video is to build systems that can in turn build other systems. Instead of sitting and managing ten AI agents by hand, you can build one overall solution that assigns tasks, coordinates the work and checks the results.

This does not mean that more agents automatically give better results. What matters is how much useful work the system actually does, how reliable it is and how little unnecessary human follow-up it needs.

The next step with AI is therefore not only about producing software faster. It is about building systems that can carry out, check and improve tasks across the whole business on their own.

And perhaps the most important question we should start asking ourselves is the following.

Why am I still a necessary part of this process, and what would it take for AI to do the job without me?

Key takeaways

• The human who writes every instruction and checks every result becomes the bottleneck as AI agents get good at programming.
• The human role has three levels, in the loop, near the loop and out of the loop. Near the loop is a natural step in between.
• An agent that is going to work on its own needs four things, namely context, model, instructions and tools.
• Ready-made products like Grok Bot and dots in ChatGPT show that agents that work on their own and are started by events already exist.
• The principles also apply outside programming, but that requires the right data, access and control.

Sources and further reading