Capital de Tokens
September 30, 2026

The map that got an agent to answer in 38 seconds

An artificial intelligence agent spent more than twenty minutes searching for an answer and ended up getting it wrong. With the same model and the same question, but with access to a relationship map, it answered correctly in thirty-eight seconds. This example, published by Uber, explains why some companies are looking

#Shorts
September 30, 2026

AI can approve its own mistakes

A less visible risk emerges here. If the same agent interprets a requirement, writes the code, and generates the tests, it can make the same mistake in all three steps. Suppose a business rule says that the discount is applied before calculating taxes. The agent understands it backwards, implements the calculation

#Shorts
September 30, 2026

A bot can be disciplined… until you give it a vague mission

Here is my least intuitive point: I do not think the sensible response is to ban all automation. A limited rule, such as maintaining a predetermined investment allocation, could be less impulsive than an investor who buys after reading every alarming headline. The bot does not get scared or need to recover

#Shorts
September 30, 2026

Enterprise AI depends on understanding how the business works

Here is my thesis: in many companies, the advantage of AI will depend less on choosing the model that writes best and more on precisely defining the world in which that model can act. What “urgent order” means. Which data is up to date. What operations exist. Who can authorize them. Those decisions seem

#Shorts