Systems, launches, and operating notes
Practical thinking on AI, automation, and the systems that replace manual work.
The boring stack that keeps AI agents shipping
The cool part of AI is the model. The shipped part is logging, retries, versioned prompts, permissions, checkpoints, and rollback. The pieces we will not run an agent without, and three failures they prevent.
A two-week rollout plan for shipping agents safely
Week 0 defines the workflow and the metric. Week 1 ships read-only with human checkpoints. Week 2 adds automation, monitoring, and rollback. A plan a small team can follow without a platform team.
The boring automation that prints time
The automation with the best return is rarely impressive. It takes one recurring decision off your plate: triage, routing, summarization, or follow-up. The hidden ROI, the four patterns, and the four-piece stack.
The three prompts that actually ship product
Most prompts optimize for answers. The ones that ship optimize for decisions and produce artifacts you can paste into a tracker. Our spec-to-tasks, critic pass, and launch copy prompts, with the templates.
Building in public without turning it into a second job
Building in public works because it creates feedback loops and trust. It fails when it becomes a performance. A lightweight system: capture raw updates while building, repurpose once, keep a rolling two-week calendar, ship first and post second.
Bourdon vs Mem0: two answers to "why doesn't my agent remember"
Both give AI agents memory that survives past one session. They disagree about where it should live and what "shared" means. A fair, source-checked comparison.
Bourdon is live: one memory for every AI you use
We launched Bourdon on Product Hunt: a shared, file-owned memory for Claude Code, Codex, Cursor, and any MCP agent. Plus the two-day hardening sprint that shipped 0.13.0.
RADLAB relaunch field note
The new RADLAB site moves the public surface into an owned system built for content, services, and future AI tools.
AI systems need operating taste
The useful AI work is rarely the model alone. It is the contract, the interface, the fallback, the evaluation loop, and the judgment around what should stay human.
Cloudflare, Vercel, and the front door
A clean website migration is mostly about the front door: DNS, redirects, certificates, headers, analytics, and the ability to roll back without panic.
Have a system that should exist?
Tell us what manual work is slowing you down. We'll respond within 24 hours.