Why Agents Stumble at Step Seven: The Simple Maths of Long Tasks

Here's a puzzle that explains a lot of agent frustration. Imagine an agent that gets each individual step right 95% of the time. That sounds excellent, and most of us would happily hire a person with that record.

Now chain the steps together. If every step has to succeed for the task to succeed, the probabilities multiply:

• 7 steps: 0.95⁷ ≈ 70% chance the whole task works.
• 10 steps: about 60%.
• 20 steps: about 36%.

Even at 99% per step, twenty steps lands around 82%. As the task gets longer, reliability quietly drains away. That's why a short demo can look magical while the same agent fumbles a real afternoon of work.

This simplified maths assumes each step fails independently and that one slip sinks the whole job. Real agents can sometimes notice a mistake and recover, which helps. Errors that snowball, where one wrong step makes the next more likely, make things worse.

How builders fight the maths

• Shorter chains. Break one long task into several small ones, each with a clear result.
• Checkpoints. After a few steps, check the work against something concrete, such as a row count, a test or a screenshot, before continuing.
• Self-checks. Ask the agent to review its own output against the goal before moving on. It won't catch everything, but it raises the odds for each step.
• Human handoffs at the risky step. Put a person where an error would be expensive, not everywhere.

There's a design lesson hiding in the numbers too. Moving one step from 95% to 99% helps far less than cutting a twenty-step chain into four five-step chains that each get checked. Structure beats polish.

So when an agent fails at step seven, it may not be "dumb." It may just be running into the arithmetic of long tasks, and you can design around that.