Writing

29 Sept 2026

Evaluating AI Agents: Building Tasks That Grade Them Fairly

Build one coding-agent task from instruction to verifier, then test it for loopholes, flakes, reward hacking, and repeatable results.

16 Sept 2026

Durable Agents: When an Agent Loop Needs a Workflow Engine

Follow a long-running agent through checkpoints, tool calls, approvals, budgets, crashes, and deploys without repeating costly or external work.

4 Aug 2026

Your Rate Limiter Will Fail Open or Closed: Choose on Purpose

When a shared counter fails, decide which requests may continue, which must stop, and how local limits, timeouts, and load shedding prevent another outage.

2 Jul 2026

How One Open Postgres Transaction Can Slow a Database

See how one open Postgres transaction can block cleanup, queue locks, grow tables, and threaten transaction ID wraparound, then learn how to detect and stop it.

10 Feb 2026

When Kafka Can Lose an Acknowledged Record

Follow one Kafka partition through four failure paths that can erase acknowledged records, then choose settings and alerts that preserve them.

researching

Evaluating Retrieval for Transcript Search

A measurement plan for transcript search: separate retrieval from generation, build a fair test set, calibrate automated judges, and compare dense, hybrid, reranked, and timestamp-aware search.

26 Mar 2026

Postgres SKIP LOCKED for Work Queues

SELECT ... FOR UPDATE SKIP LOCKED can support a modest Postgres work queue: workers claim rows without blocking, while leases, fencing, retries, and indexes provide the recovery behavior.