Agent design

Agents need boring guardrails

The impressive demo is an agent doing five steps by itself. The production version is knowing which sixth step it must never take.

~5 minAI
scroll
01 · the shape

Autonomy multiplies blast radius

An agent is just software with a loop, tools, memory, and permission to keep going. That loop is powerful. It also means a bad assumption can repeat until it has touched files, calendars, customers, or money.

agent · task loop

The useful signal is usually already there.

guardrails armed

Hit the button and watch the hidden AI surface area light up.

02 · the move

Design the stop signs first

Before adding tools, write down what the agent may read, what it may change, how much it may spend, when it must ask, and where every action is logged.

agent/permissions.json

A small permission file is more useful than a heroic prompt.

03 · the check

Simulate failure before launch

Run the agent against fake data with fake tools and real limits. The goal is not to prove it succeeds. The goal is to see how it behaves when the task is ambiguous, stale, impossible, or risky.

AI launch switchboard
sloppy
shippable0
fragile: the model is still freelancing without enough checks

Flip the switches and watch the story turn into a tiny operating model.

0human approval gate for irreversible work
0percent of tool calls logged
0silent writes to production systems
04 · the takeaway

The guardrails are the product

Users do not just want the agent to act. They want to trust it while it acts. That trust comes from limits, receipts, and graceful refusal.

source trail