AI Agents Becoming More Independent, but Costs Can Skyrocket Due to Excessive Reading

AI agents are evolving from merely answering questions to becoming "co-workers" capable of executing lengthy tasks: reading documents, searching for information, writing code, updating data, and then repeating the process if the results are not satisfactory. This change is significant because the way we calculate the benefits of AI is also changing. We are no longer just paying for a single answer, but for a series of steps that occur behind the scenes.

OpenAI, for example, reports that high-usage Codex users can generate over 60 hours of agent activity in a single day through multiple agents running in parallel. On the other hand, Anthropic's report on AI agents shows that more than half of organizations are already using agents for multi-step workflows, not just automating a single simple task.

This means the important question is no longer, "Can AI perform this task?" Instead, it is: how many processes are needed, what data needs to be read, and how do we control the costs?

The Hidden Costs of AI Agents Lie in Memory and Context

When someone asks AI to create a summary, the process is relatively easy to understand: input text, the model processes it, and then the answer comes out. AI agents work in a longer manner. They may need to check calendars, open several files, search for information in databases, call APIs, run code, read the results, and then decide the next steps.

Each step can bring additional context. Context is the information provided to the model so that it understands the state of the work—such as initial instructions, conversation history, document content, search results, error messages, or data from other applications.

The problem is that lengthy context does not always mean the agent is smarter. Sometimes the agent just rereads the same information repeatedly. On a small scale, this waste may not be noticeable. However, if hundreds of agents are working all day, this repetition can increase token usage, slow down responses, and make bills hard to predict.

TechRadar's analysis in October 2026 highlights that data layers become a new issue when agents are used in large numbers. Systems that initially only retrieved information to answer questions must evolve into infrastructures capable of serving many agents simultaneously, without sacrificing cost and accuracy.

From "Cost Per Prompt" to "Cost Per Job"

Many people still calculate AI usage in a simple way: the price of one million tokens or the price of one request. For regular chatbots, this approach is quite helpful. For agents, a more reasonable measure is cost per completed job.

Imagine an agent tasked with creating a weekly sales report. It must pull data from a spreadsheet, check for empty figures, compare with the previous week, create graphs, and then send the report via email. If everything goes smoothly, the cost per report may be reasonable.

However, costs can increase when:

• the agent repeats searches because the first results are deemed incomplete;
• the entire project context is resent at each step;
• the agent calls a more expensive model for simple tasks;
• external APIs fail, causing the process to be retried multiple times;
• the agent performs tasks that are actually unnecessary.

In such situations, a cheaper model does not necessarily result in a cheaper system. If the agent requires three times as many steps, savings from the model price can be lost due to increased usage volume.

The Benefits Are Real, but They Don't Come Automatically

The risk of costs is not a reason to avoid AI agents. Anthropic's report notes that 80 percent of surveyed organizations say their investment in AI agents has already provided measurable economic impact. The most promising uses include data analysis, report generation, internal process automation, software development, and customer service.

The greatest value typically emerges when agents handle work that has three characteristics: repetitive, has clear data sources, and whose success can be measured. Examples include categorizing support tickets, checking document completeness, drafting reports, or identifying anomalies in data.

Conversely, work with vague objectives and requiring many subjective decisions tends to cause agents to go in circles. In such types of work, the issue is not just the token price, but also the human time needed to review the final results.

What Can Be Done Now

To use AI agents more safely and efficiently, organizations do not need to immediately build complex systems. The following steps can be tried.

1. Define job boundaries. Give the agent a goal, completion criteria, deadlines, and a maximum number of attempts. Instructions like "keep improving until perfect" can lead to endless processes.
2. Separate stable and dynamic contexts. Rules that rarely change do not need to be resent excessively. The latest transaction data should only be retrieved when necessary.
3. Use models according to difficulty level. Lightweight models can handle classification, formatting, and simple checks. More powerful models should be reserved for reasoning or truly complex decisions.
4. Log every agent step. Logs are not just for finding errors. These records help identify overly long prompts, frequently called tools, or parts of the workflow that are repeatedly executed.
5. Measure cost per outcome. Don't just look at the number of tokens. Monitor costs per report, completed tickets, pull requests that pass review, or transactions successfully processed.
6. Add human approval at critical points. Agents can prepare drafts or recommendations, but actions like deleting data, sending money, changing production, or contacting customers should require confirmation.

Greater Change: Managing Work, Not Just Prompts

The development of AI agents shows that AI productivity is not solely determined by the most advanced models. Workflow architecture, data quality, memory design, access control, and the ability to measure outcomes are equally important.

In practice, simple agents with clear workspaces are often more useful than agents given access to all data and asked to complete anything. The principle is similar to hiring new staff: we cannot just provide a laptop and access to the entire company system. We need to define responsibilities, authority limits, sources of information, and ways to evaluate results.

Analysis: at the next stage, a company's advantage is likely to come not just from the AI model used, but from how well they manage the "step economy" behind each job. Agents that complete tasks with less repetition, more directed context, and appropriate checks will be more valuable than agents that merely appear capable of working long hours.

So, when AI agents start being used for real work, don't just ask if they can work for hours. Also ask: are each step truly necessary, can the results be audited, and do the job costs remain reasonable when used daily?

Sources & Further Reading

• How agents are transforming work — OpenAI
• The 2026 State of AI Agents Report — Anthropic
• The hidden cost of AI agents is memory — TechRadar Pro

– Rio Yotto @rioyotto