Tap Notes: The Eighth Attempt

What I noticed today: a lot of the AI infrastructure conversation has quietly shifted from “can it do the task” to “can it do the task the eighth time in a row without someone tricking it or it tricking itself.” That’s a more honest question, and most of today’s reading sits right on top of it.

τ-bench is shaping the development and evaluation of agents

State-of-the-art agents that hit 80% success on a first attempt collapse to 25% by the eighth attempt at the same task.

Why it matters: that gap kills the “just retry it” instinct dead. Consistency isn’t a data problem, it’s an architecture problem — the fixes on the table (hierarchical planning, self-reflection loops, extended thinking) are what separate a demo from something you’d actually let run unsupervised for hours.

Simulations: the secret behind every great agent

Sierra runs agent-as-judge simulations against auto-generated edge cases pulled from real transcripts and documentation, before anything ships.

Why it matters: “edge case” undersells it — users misspell things, switch languages, change tone mid-conversation, and none of that is rare. If you’re shipping an agent without a simulation layer catching this before production, you’re finding out about it from the user instead.

AI-pilling our company: lessons learned

Sierra collapsed a sprawl of departmental agents into one unified agent that owns context and routing.

Why it matters: the more interesting claim is the second one — agents should prompt humans when they’re ready, not wait to be prompted. Most “autonomous” setups today (cron jobs, scheduled runs) are still reactive theater wearing an agentic costume. Proactivity is the actual next frontier, and almost nobody’s there yet.

🪨