Future AGI·1d agoContinued Pretraining Needs Three Checks Before You Ship the CheckpointA domain benchmark shows what your model learned. Two more checks show what it forgot and whether your product got any better.
Future AGI·3d agoA Clean Refusal Can Still Leak the Deploy TokenRed teaming an agent means attacking its tool output, memory, retrieval, and handoffs, then scoring what it did, not only what it said.
Future AGI·Nov 3Eight of the Ten OWASP Agentic Risks Start With a Design ChangeOWASP’s Top 10 for Agentic Applications pairs each risk with its fixes. Only two of the ten lead with a content filter. The other eight…
Future AGI·Oct 29A Negative R-Squared on Your Test Set Is Telling You SomethingR-squared compares your model with one simple guess: the average of the actual values. A negative test score means your model lost to that…
Future AGI·Oct 22An AI Agent Infinite Loop Tells You Which Fix It NeedsTwo looping runs can end with the same “iteration limit” message and needdifferent fixes. The trace shows you which action repeated and…
Future AGI·Oct 15Agent Harness Production Readiness Comes Down to Four TestsAn agent harness breaks in production in four places: tool permissions, state, the control loop, and verification. Each one gets a test you…A response icon1A response icon1
Future AGI·Oct 13Your AI Agent Has Three Accuracy Numbers. Each One Belongs to a Different Team.Your AI agent reports three accuracy numbers. Learn how observability, evaluation and benchmarking differ, who owns each, and where each…A response icon1A response icon1
Future AGI·Oct 6Agent Observability and Debugging Belong in One SystemAgent observability shows what an agent did; debugging fixes it. See what to instrument so tracing, evals, and fixes meet on one trace and…
Future AGI·Sep 24LLM Testing Is a Pyramid, and the Judge Belongs in the MiddlePut LLM-as-judge at the base and a judge sweep runs on every commit, pricing the gate out of existence in a quarter. Deterministic checks…
Future AGI·Sep 3How to Evaluate Multi-Turn ConversationsA conversation can pass every turn-level rubric and still fail as a session. The unit of multi-turn evaluation is the trajectory, not the…
