Notes

MAY 12, 2026

7 min read

Building agentic workflows that actually recover from failure

Most agent demos work on the happy path. Real agentic systems live or die by how they handle the failure path — the calls that error, the tools that lie, the steps that get half-done.

APR 24, 2026

6 min read

The agent harness is the real product

The model gets the headlines. The harness — tools, sandbox, permissions, observability — is what actually determines whether an agent ships or quietly gets shelved.

APR 02, 2026

5 min read

Why interpretability matters for the people shipping AI

Interpretability is often framed as a research concern. For builders, it is the only honest answer to the question every stakeholder eventually asks: why did the model do that?

MAR 14, 2026

8 min read

Agentic search vs RAG: when reasoning beats retrieval

RAG retrieves once and prays. Agentic search reasons over results, refines its queries, and follows the lead. They solve different problems — and people pick the wrong one all the time.

FEB 18, 2026

7 min read

AI products need operational safety, not just model safety

Most reliability work stops at the prompt or the model card. Production AI needs operational safety: permissions, logs, rollback paths, and explicit human ownership.

JAN 30, 2026

6 min read

Reasoning models still need someone to check their work

Chain-of-thought made models look smarter. Verification — by another model, a tool, or a human — is what actually keeps them honest on the problems that matter.