I Do Run AI Agents Overnight. Here's What Actually Matters.
Boundaries, supervision, evidence, and why unattended execution is not the same thing as unbounded autonomy.
Petie Clark — Applied AI Systems
Applied AI systems that survive contact with the business.
Field notes on agent orchestration, evaluation, authority boundaries, enterprise integration, deployment, and production operations.
Get new posts in your inbox
No spam. No marketing. Just the writing.
/start-here
How I think about governed systems
Why I Prefer One Clear Source of Truth
Why tools can multiply while authority, state, and accountability still need one dependable home.
Why I Write Down the Rollback Before the Change
A deployment is not ready merely because the forward path looks convincing.
The Case for a Smaller Blast Radius
Production systems become trustworthy through bounded change, not theatrical confidence.
AI Can Compress the Work Without Owning the Judgment
The useful AI review packet shortens evidence gathering while making the reviewer’s corrections and acceptance easier to see.
4 min read
Read →
The Handoff Is Part of the Architecture
A system has not been operationally accepted until someone else can use its maintained material to explain it and safely perform ordinary work.
3 min read
Read →
Technical Debt Is Sometimes Just Unmade Decisions
Decision memo: remove competing service routes by choosing one authoritative path and making the fallback prove its limits.
4 min read
Read →
Model Evaluation Is Mostly About Choosing Better Failure
My model notes became useful after I stopped looking for one winner and started recording which mistakes each workflow could survive.
4 min read
Read →
A Workbench Beats a Platform More Often Than Vendors Admit
A useful internal surface can unify a job without trying to own every system behind it.
4 min read
Read →
The Interface Between Humans and Automation Is an Authority Model
Automation becomes predictable when permission is part of the interface, not an assumption hidden behind it.
3 min read
Read →
I Stopped Counting Activity and Started Counting Accepted Work
The useful measure of automated work begins after output exists: was it verified, accepted, and produced with less steering?
4 min read
Read →
Release Gates Are How Fast Teams Stay Fast
A release candidate moved quickly because its identity stayed fixed and every available gate left evidence.
4 min read
Read →
The First Job of an AI Agent Is to Preserve Context
An agent can stop, fail, or be replaced without losing the work if state lives in evidence-bearing artifacts.
4 min read
Read →
Good Architecture Makes Failure Boring
The architecture earns its keep when a failure stays inside a boundary, lands in a named mode, and leaves us a sane way back.
4 min read
Read →