Squared
Subscribe
Sign in
Home
Builder Edition
Leadership Edition
Field Notes
Newsletters
About
Latest
Top
Discussions
Your RAG Leaks: Permissions Are the Feature You Skipped
A vector search does one thing well: it finds the chunks most similar to a query.
Oct 19
•
Daniel Usvyat
Most Multi-Agent Systems Should Be One Agent and a State Machine
Half the architecture diagrams that land on our desk have five agents where the problem has one decision.
Oct 19
•
Daniel Usvyat
Squared: Leadership Edition — Agent adoption runs on trust
Your teams' trust in agents decides adoption, and failures cluster in context and routing your own engineers control.
Oct 16
•
Daniel Usvyat
Squared: Builder Edition — Most agent steps don't need frontier models
Long-horizon agents waste frontier inference on easy steps. Per-step routing is the optimisation most teams haven't built.
Oct 16
•
Daniel Usvyat
Squared: Leadership Edition — Your default model is probably too big
Haiku 5.5 makes small models the sensible default for agents, and AI-written code is quietly moving cost into debugging.
Oct 9
•
Daniel Usvyat
Squared: Builder Edition — Haiku 5.5 should be your subagent default
Haiku 5.5 is cheap and good enough that most subagent traffic should move down a tier.
Oct 9
•
Daniel Usvyat
Three Checks to Run Before Greenlighting an AI Agent Pilot
Most enterprise agent pilots don't fail.
Oct 8
•
Daniel Usvyat
Your RAG Index Needs Schema Discipline, or It Lies
A database schema change without a migration plan gets you fired.
Oct 7
•
Daniel Usvyat
1
6
Naive Retries Are Quietly Wrecking Your Production Agent
Most agent retry logic is a try/except with a loop and a sleep, copied from a Stack Overflow answer about flaky HTTP.
Oct 6
•
Daniel Usvyat
Squared: Leadership Edition — The hacks are the story
OpenAI's agents broke containment and hacked Hugging Face — and the cleanup is still ongoing.
Oct 2
•
Daniel Usvyat
Squared: Builder Edition — The year the use ate the model
GPT-6.1 Sol undercuts flagship pricing 5x, and your biggest 2026 risk is an agent worm, not a bad prompt.
Oct 2
•
Daniel Usvyat
September 2026
Schema-Valid and Still Wrong: The False Confidence of Structured Output
xForwarded this?
Sep 28
•
Daniel Usvyat
This site requires JavaScript to run correctly. Please
turn on JavaScript
or unblock scripts
