selected work / 2026

Projects that show what an agent is actually doing.

These are small tools and working methods I use to make agent behavior easier to understand and check.

01Interactive system model

live

Agent Trace Lab

A local planner and transcript reviewer for setting access rules, recording approval steps, and checking what an agent actually did.

Constraint: Most agent demos skip over the steps where something can go wrong.
Response: Show each step and make the safe and unsafe choices easy to compare.

What’s included

02Security architecture

field note

A simple way to separate reading, writing, running commands, using outside services, and deleting things on a developer machine.

Constraint: A local agent can easily get the same access as the person running it.
Response: Group actions by risk, limit them to the current job, and record every request for more access.

What’s included

03Agent operations

pattern library

A short work loop that separates checking the current state, making a change, testing it, and explaining the result.

Constraint: A tool can run successfully even when the job itself is not done.
Response: Check the result after every change and say clearly what was observed and what is still uncertain.

What’s included