My agent
workbench

Making things with AI.
Keeping enough evidence to trust them.

John Lee

AI & Ales · Jamestown Frontier Society
October 15, 2026 · Billsburg Brewery

The tools around my agent

Codex

A working partner in my projects, files and terminal

Skills + Markdown

Reusable instructions and decisions that survive a chat

Git + worktrees

Separate experiments, inspect changes, recover a state

Pixi + DataLad

Recover environments and connect data to the work

Keyhatch

Selective credential access for agents

Approve selected credentials once.
Allow repeated retrieval within a limited window.

Native macOS app + local command-line client

Expiry stops future retrieval. It cannot recall a disclosed secret.

Experimental build

The Software Engineer Reborn

I used an interactive preparation space to work through the ideas.

What did we just make?

A research object includes
the result and what it takes
to inspect or rerun it.

STAMPED asks about the package:
its history, its parts, its environment,
and how someone else can use it.

Orinoco Lite

Connected research information

People, projects and outputs
share explicit relationships.

An editor proposes changes.
GitHub review preserves the history.
The records feed a website and graph.

STAMPED in Practice

Skills that improve through
scientific work

Study record on Zenodo

Next: run, assess, improve, repeat

Compare four assistance conditions.
Test quality, effort and consistency.

Research workshop in development.
Transfer remains to be tested.

What would
convince you?

What evidence do you need before you trust
something an agent made?

I’m exploring how to make that evidence easier to keep and inspect.