What the agent walks into

When ChatGPT showed up, the surprise was not that it was smart. It was that it talked. For the first time, software felt conversational. That pulled AI out of demos and into daily work.

In Brazil, it landed even harder because we already live inside messaging. WhatsApp is where we schedule, buy, sell, solve problems, and run parts of companies. It did not take long for "AI" to look like a bot in a chat.

Chat is a natural place to state an intention. An agent then has to work with the world behind it. A sales lead belongs to a company record, an approval, and a conversation. A move belongs to dates, addresses, reservations, and the people travelling.

That is a different expectation of software. It has to find the relevant state and decide which sources can be trusted. When the evidence leaves a conflict unresolved, it should ask before acting on that uncertainty.

Most of what matters never gets typed into a system. It lives in calls, messages, side conversations, and decisions made out loud. That is how work happens, but most software was never designed to capture it.

The problems are ordinary: a stale field in the CRM, two definitions of the same metric, a meeting that never made it into any system, or the same customer represented three ways across billing, support, and analytics. An agent can follow its instructions precisely and still act on an incomplete version of reality.

OpenAI's October 2025 company knowledge launch addresses part of this problem by searching connected work apps and citing the sources behind an answer. That gives the agent more evidence to work with. An agent can help match customer records and expose conflicting definitions. The people responsible for reporting still need to settle which definition of revenue should govern the decision.

People patch over this mess all day without noticing. Software follows what the systems say, often with more confidence than the data deserves. I have seen teams explain failures with "That field was never documented" or "That meeting was not in the CRM, so the agent could not know." The model gets blamed for the company it was dropped into.

Deploy the same agent in two companies and you get different results because each company defines reality differently. What qualifies as enterprise? Which system owns revenue? How is churn calculated? Which number do people trust?

For a personal agent organizing a move, the relevant information might be in messages, a calendar, and delivery accounts. Suppose the moving date changed in a conversation, but the calendar still has the old date. The saved delivery address is also about to become wrong. Access to all three would let the agent find the disagreement; it would still have to establish when to use each address.

A company can assign someone to maintain those records. At home, I would expect the agent to help me resolve the conflict without making me maintain another database. A clear confirmation in the messages could establish the new date. If the change is still only a suggestion, it could ask which date holds and keep that answer attached to the move. I should not have to explain the correction separately to every service it uses.

Even with the facts settled, two people can need different answers. One wants a headline; another wants a drill-down. One wants something to act on now; another wants every caveat. Those preferences change the presentation, not the facts or the authority behind them.

I used to compress all of this into the word memory: identity, definitions, history, context, and the system learning how someone works. I no longer think the word can carry that much.

Some of these things define the world the agent operates in. Some describe its current state and what happened before. Some govern what the agent may do. Others develop through repeated interaction with a person. Grouping them together hides the work each one has to do.

Once agents take action, the gaps become harder to ignore. A delivery goes to the wrong address, a discount reaches the wrong account, or a message goes out before someone agreed to send it. The missing context affects other people before anyone gets to correct the answer.

I would want the agent to help make sense of that environment as it works. It could connect records, surface a disagreement, and keep a correction where the next task can use it. People should not have to resolve every inconsistency before getting anything useful from it.