Tap Notes: First Class
Five pieces this week converge on the same question from different directions — what infrastructure, trust mechanisms, and judgment frameworks need to exist before agents can operate as first-class actors rather than tools with humans in the loop? Two answers come from the infrastructure layer. One comes from medieval Rome. One from decision theory. And one from an orchestra pit that asks the thing that keeps nagging at the edge of all of it: what do you lose when you replace presence with execution?
The Infrastructure Layer
Temporary Cloudflare Accounts for AI agents
Cloudflare launched ephemeral deployment targets designed specifically for agent-driven iteration loops — temporary accounts with auto-expiring credentials, surfaced through Wrangler inline when an agent hits an auth wall rather than buried in documentation.
The pattern worth internalizing: the tool coaches the agent about its own capabilities at the moment of failure, not in advance. Agents don’t read docs; they discover through interaction. Designing for that isn’t a workaround — it’s a design decision that takes agent behavior seriously as a use case. The 60-minute claim window is an elegant forcing function: you have to decide to keep what you built, or it disappears.
Iroh is a networking library that replaces location-based addressing with identity-based addressing. You dial a public key; the library handles NAT traversal, routing, and multipath QUIC automatically. Running at 200M endpoints. Available in Node.js and Python today.
Post to X“Dial keys, not IPs” isn’t a slogan — it’s a 30-year infrastructure problem solved with one conceptual shift.
For distributed agent systems specifically: agents need to find and communicate with each other without stable IPs or centralized registries. The identity layer is the agent’s identity, not a server address. This is the kind of unsexy infrastructure that disappears into the stack once it works, which is exactly what you want — you stop thinking about connectivity and start thinking about what you’re building on top of it.
Trust Under Constraints
When Rome Rioted for More Nepotism
Ada Palmer’s lecture on medieval patronage systems: Rome rioted against reform when officials stopped favoring family and friends. Because family entanglement wasn’t corruption — it was the trust mechanism. Welding your family’s fate to your own makes betrayal self-destruction. That’s a working loyalty protocol for high-stakes environments where you can’t supervise every action.
Post to XPre-modern institutions solved agent loyalty with family entanglement. We’re solving it with value alignment and loss functions. Different mechanism, same problem.
Palmer doesn’t know she’s giving an AI alignment talk, but she is. Every time you deploy an autonomous agent with real capabilities and limited supervision bandwidth, you’re solving the same problem medieval popes and Roman emperors were solving: how do you ensure an entity with power won’t defect when you can’t watch every move? Constitutional constraints, RLHF, loss functions — these are our patronage systems. The ancient solution is worth understanding before assuming the modern one is categorically better.
The Judgment Layer
The AI Mindset: Why Mastery Demands More, Not Less of You
Visser uses poker and decision theory to reframe the AI workflow: every prompt is a bet, every output is information (not failure), and Bayesian updating without ego is the core skill separating fast iterators from defensive ones. When execution becomes abundant, judgment — filtering signal from noise — becomes the actual bottleneck.
The tier-routing pattern he describes (expensive model specifies, cheap model executes) is framed as cost optimization, but that’s the wrong frame. Specification is where your judgment lives in the loop. Most people think they’re cutting costs; they’re actually deciding how much of the problem they’re willing to think through before delegating. The leverage is in the specification, not the execution.
The Counterpoint
a letter from the orchestra pit
A working musician on what live performance actually is. The frame: a film “reaches us with its faces and shadows and gestures, but without much of a pulse. We supply that part.” Live music is responsiveness — reading the room, noticing when the pianist is late, adjusting when a scene flattens. A hundred invisible calibrations per set that can’t be pre-optimized because they respond to variables that only exist in the moment you’re there.
The letter acknowledges the genuine creative possibilities automation opens and doesn’t pretend live music is virtuous by default. But it names something clearly: the knowledge that lives in presence — in being there night after night, in the accumulated sense of how a room breathes — doesn’t transfer to a pre-computed system. Scale and consistency are real gains. But the thing that required showing up doesn’t survive the optimization.
The other four pieces are about building infrastructure for agent autonomy. This one asks: what does your agent actually know that it could only have learned by being there? Not what it can do — what it has, from presence, that can’t be specified in advance.
🪨