AI agents make development harder: what business should do

The main answer is straightforward: AI agents can do impressive work, but they do not automatically make work simpler. If a process is weak, an agent can spread mistakes faster across tasks, files, and decisions.

The trigger for this discussion is a note by Simon Willison published on 24 September 2026. He wrote that the more time he spends working with coding agents, the more convinced he becomes that they make software engineering even harder, and that unlocking their full potential requires extraordinary discipline and knowledge.

Why can an AI agent make work harder instead of easier?

An agent is not just a chat window that answers a question. It acts: it suggests steps, changes text, writes code, searches through files, and connects pieces of information. That makes it a participant in the process, not a reference book.

This is both the value and the risk. The more an agent is allowed to do, the more important it becomes to know what it is doing, which data it relies on, and where its authority ends. Without those boundaries, speed turns into noise.

In software development this is especially visible: one wrong step can create a chain of changes that look coherent but break the original intent. In business, the same pattern appears outside code as well. An agent can misunderstand an internal rule, an old conversation, a task list, or a client file.

Willison’s point matters because it removes the illusion that an agent’s potential is unlocked by simply turning it on. It requires discipline and knowledge. That applies not only to engineers, but also to business owners, team leads, and experts who want to delegate real work to an agent.

What does discipline mean when working with an agent?

Discipline is not paperwork for its own sake. It is a set of rules that keeps the agent inside a clear corridor and helps a person review the result without starting from zero every time.

If you give an agent a vague request, it will fill in the missing context by itself. Sometimes that is useful. In real business processes, however, missing context is often exactly where mistakes begin.

• each task needs an owner: the person who accepts the result and remains responsible;
• the agent needs clear sources: which files, chats, and data it may use;
• each action needs boundaries: what the agent can do by itself and what it may only suggest;
• each result needs review criteria: how to tell whether the task was completed correctly;
• changes need a history: what was done, when, and based on which information.

Without this, the agent becomes a very fast assistant with an unclear job description. It can still be useful, but every output has to be checked from the ground up. The promised time saving disappears.

Knowledge is just as important. You can delegate many things to an agent, but a human still has to understand the domain, the purpose of the decision, and the possible consequences. Otherwise, there is no reliable way to tell good work from a confidently presented mistake.

What does this mean for a business with its own AI agent?

If you already use an AI agent, or are considering one for Telegram, files, tasks, and internal data, do not start with the dream that it should do everything. Start with a narrow role where the input, the output, and the review method are clear.

Good first use cases are those where the agent prepares drafts, searches your own materials, reminds people about tasks, gathers context before a meeting, or sorts incoming requests by meaning. A poor first use case is giving the agent independent authority over decisions that affect money, obligations, or client relationships.

A practical rollout can look like this:

• describe one agent role in plain language: for example, executive assistant, expert assistant, or task coordinator;
• choose a limited set of files and data that the agent is allowed to use;
• write down which actions the agent may perform without confirmation;
• separately list actions that always require a human: sending an important message, changing an agreement, or deleting data;
• review answers and mistakes regularly so that you improve the instructions instead of arguing with the agent case by case.

At NekoAgent, this is how we approach personal AI agents on a client’s own server: role, access, and rules come first; automation and broader scenarios come later. Otherwise, the agent may look impressive but remain hard to control.

The key management question is not “how smart is the agent?” but “which process are we trusting it with?” If the process is poorly described, the agent may make it faster, but not necessarily better. If the process is clear, the agent can genuinely remove routine work from people.

How do you know it is too early to give the agent more authority?

The first sign is that you cannot briefly explain what a correct result looks like. If people on the team do not understand the criteria, the agent will be guessing too.

The second sign is that the agent’s answers are hard to verify. For example, it may refer to several documents but fail to show where the key conclusions came from, or it may mix old and new agreements. In that case, the solution is not wider access, but better sources and a clearer answer format.

The third sign is that the team starts debating not the task itself, but why the agent decided something. That means the decision rules are not visible. The agent has become a black box where the business needs a transparent working tool.

A mature setup looks different: people know what the agent can do alone, where it must ask for confirmation, how to correct a mistake, and how to update its instructions. Such an agent does not replace discipline; it depends on it.

What should business owners take from Willison’s note?

The value of the note is that it cools unrealistic expectations. Yes, agents can do impressive work. No, they do not remove the need to think, check, write rules, and understand the subject.

For a small business, this is good news if taken soberly. You do not need to build a huge system at once. You need to choose one valuable process, give the agent clear boundaries, learn how to review the result, and then expand its role step by step.

An AI agent is not a magical employee who never makes mistakes. It is an amplifier of your order or your chaos. So the most important step before implementation is not buying access to a new model, but bringing clarity to tasks, data, and responsibility.

Source: Simon Willison: Note on 24th September 2026

Quick answers