What autonomy costs
By the end: Name the three costs autonomy adds and estimate them for a proposed agent before building it.
Autonomy is not free, and the three prices it charges are predictable enough that you can estimate them on a napkin before writing any code.
The first is money. Every pass through the loop re-sends the whole conversation, so cost grows faster than the number of steps — roughly with the square of it, since each step both adds tokens and re-reads all the previous ones. A ten-step agent is not ten times a one-step call; it is considerably worse unless you manage context deliberately.
The second is time. Each pass is a full model call, and they are sequential by nature: the agent cannot decide step four until it has seen the result of step three. An agent that takes twelve steps will feel slow no matter how fast each call is, which changes what interfaces you can put it behind.
The third, and the one that surprises people, is compounding error. If each step is right ninety-five per cent of the time, twenty steps in sequence gives you roughly a one in three chance of a clean run. Reliability per step has to be very high before long chains are usable, which is why feedback that lets an agent notice and correct its own mistakes matters so much.
None of this is an argument against agents. It is an argument for knowing what you are buying. When the task genuinely has an unknowable shape, an agent that costs a dollar and succeeds is far cheaper than a person spending twenty minutes. The mistake is paying agent prices for workflow problems.
End of module 1