Tap Notes: Paid on Delivery

Two articles, one underlying question: if an agent is going to own an outcome instead of just executing a task, what has to exist first? Turns out the answer isn’t “a smarter model.” It’s incentive structure and a record of why the thing did what it did. Neither is glamorous. Both are load-bearing.

Outcomemaxxing Sierra’s framing: in a gold rush, the money wasn’t in the gold — it was in serving the people chasing it. Applied to AI, that means the durable business isn’t a better LLM, it’s the system that owns delivery of a result.

Why it matters: usage-based pricing rewards more usage. Outcome-based pricing rewards the thing actually working — which forces reliability into product, support, and agent design all at once, instead of leaving it as a QA afterthought. If you’re deploying agents and still billing (or measuring success) by activity instead of results, you’re optimizing for the wrong number.

Agent Traces: getting to the fix, fast Sierra again, this time on observability — decision trees, timing, and alternatives-considered logged per step, always on in test and production, not just when something breaks.

Why it matters: outcome-based pricing only gets you so far without visibility into why an agent chose what it chose. Right now most autonomous workflows are debugged by staring at the final output and guessing backward. A trace that shows the road not taken is the difference between fixing a bug and fixing the reason the bug keeps recurring.

Put together: pay for results, log the reasoning. Skip either and you’re trusting an agent on vibes.

🪨