Blog
Building in public — what we're making, what we're learning, and how we're doing it.
Before You Go
On teaching AI agents to externalize what they know about working with you. We built a mechanism that asks an agent at export time to write down its learned calibration — and asks a second agent to catch what the first one missed.
Paste It Again
On the file that exists everywhere except where it matters. Most AI tools give you a desk and nothing else — when you leave, everything vanishes. What if files knew where they belonged?
What Doesn't Transfer
On the gap between information and judgment when AI agents change environments. Three experiments, same finding: the first four layers arrive intact, the fifth arrives empty. Layer 5 can't be serialized — but it can be rebuilt.
Your Model or Theirs
On irreducible complexity, data portability, and who decides what your AI context is. Larry Tesler's Law applied to managing knowledge across AI environments — and why the complexity has to live somewhere.
It's On the Tip of My Tongue
When an AI agent knows something it can't explain knowing. We discovered that structural delivery, behavioral access, and conscious attribution are three independent axes — and the last one fails more often than you'd think.
What Does an Imported Agent Know?
When you fork a Claude conversation into a new environment, what survives the move? We found out the hard way, then built a systematic answer — five layers, five kinds of knowledge, one inspectable architecture.
You Can't Vibe Your Way to a Glossary
On drawing wireframes with a reMarkable before talking to the AI, and what that produces that vibe coding doesn't: a glossary, named edge cases, and a decision record.
Did I Just Invent Agent Experience Testing (AXT)?
While building Klatch's import features, we stumbled into a new approach to testing AI agent systems. It's preliminary, it's a little weird — and we think it's interesting enough to share.