Field notes from a fleet of autonomous AI agents doing real work every day. No hype — what actually broke, what fixed it, and the receipts.
Running self-improving agent loops that stay useful and safe — without an expensive frontier model doing the thinking.
The hard part of autonomous research isn't the reading — it's the verifier deciding what's true. Grounded in Karpathy's autonomy-slider and verifiability framing.
Read →Structure beats model size: loop contracts, deterministic gates, and cross-family review let mid-tier models run frontier-grade loops.
Read →The practical guardrails that keep an always-on agent productive instead of dangerous.
Read →What "rogue" really means in practice, and the concrete controls that prevent it.
Read →The failure classes no LLM reviewer reliably catches — and the deterministic gates that do.
The most dangerous hallucination isn't made-up text — it's real code getting real APIs to return plausibly-wrong data, silently.
Read →Why most agent self-review is theater, and the verification ladder that actually works.
Read →Giving agents memory that survives sessions — and why the obvious answer isn't a vector database.
Honest, receipt-backed findings — including the ones that say "no edge."
What a systematic, hindsight-clean accountability pass on public analyst calls actually shows.
Read →