The Constraint Problem
Why agents excel at coding but fail at project planning. The constraint problem, comprehensive context, and the Socratic move that closes the loop.
Bart Kolendowski
Researcher & Co-founder
the engineering around the agents
I'm a researcher working on bioinformatics with AI agents in Canada. I also co-founded HoppeSyler Scientific — we make reagents and ship them to life science labs across Canada and the USA.
HoppeSyler began as a conventional reagent business but in 2022 (when ChatGPT was still just "GPT") I saw the writing on the wall and started the process of rebuilding and organizing everything around the inevitable rise of AI. Today, 3.587123 years later, personalized AI agents permeate the entire structure — running every aspect of the business that doesn't require physical labor.
Three years in, the surprise has been how little of the work is the agents themselves. Most of it has been the engineering around them — making sure they fail safely, escalate when they should, and don't ship something they shouldn't.
This is where I write about what I've learned from doing it — the patterns that hold up under real stakes, the places where agent leverage breaks down, and the operator-side discipline that emerges when models can produce more than you can verify.
recently writing
essays from real-stakes agent work
Language Models: Patterns of Engagement
Five asymmetries between managing people and managing language models. Three anti-patterns. One meta-principle.
Agentic principles
patterns that have held up under real stakes
- Generate, don't iterate.
Volume-plus-selection beats single-pass refinement when generation is free and evaluation is the bottleneck. Five variants, pick the best, refine the parts still weak. The operator surveys a landscape, not climbs one hill.
- Radical candor has zero cost.
A model has no ego, no morale, no political standing. Every hedge word inherited from human collaboration is a tax with no return. Prompts should read as specifications, not messages to a coworker.
- Fresh sessions are free red teams.
One context builds, another attacks. Adversarial review used to be infrastructure only large teams could afford. Now it's a second conversation window — and almost no operator is using it.
- Context documents are infrastructure.
In a model-augmented workflow, written context becomes the memory substrate every session loads from. Maintain it the way you maintain a database schema — versioned, refactored as the work evolves.
→ full essay: patterns of engagement
building
tools running the business right now
-
▸operator-dashboardWhere agents run the back office. Inbox triage, inventory sync, escalations, change feed — work that normally requires headcount.live
-
▸grant-scannerAn always-on prospector. Reads federal grant databases, qualifies bioinformatics fits, feeds the sales pipeline.live
-
▸paper-radarAccount intelligence at scale. Connects fresh publications to active customers — surfaces what each lab is doing and what they'll need next.beta
-
▸agent-osThe substrate the agents run on. Projects, weekly cycle, journals, communications — all addressable, all auditable, all under version control.live
code, last year
contributions, last 365 days
52 wks ago
today
agent-augmented output. commits is now easy; the real work is in deciding what to ship.
reading
on the desk
- currentlyDeep Work: Rules for Focused Success in a Distracted World
Reading now.
- referenceThinking in Systems: A Primer
Read. Pulled off the shelf before every new problem.
- readZero to One: Notes on Startups, or How to Build the Future
Read once. Distribution and the 10x rule were the parts that stuck.
- revisitsThe Beginning of Infinity
The kind of book where rereading is the point.