Writing · long-form
Long-form on agents, evals, internet-native building, and the unsexy plumbing that turns a demo into a system.
Most teams already use ChatGPT. Most still wait the same number of days for a quote, a reply or a decision. AI is speeding up one step. The handoffs between steps haven't moved. Rewrite one end-to-end process first, give it one owner, then put AI inside it.
AI isn't mainly wiping out entry-level jobs. It's taking over entry-level tasks, like research, summaries and first drafts, and those tasks were how juniors learned the job. Young workers in AI-exposed jobs are already being hired less, while the roles that remain ask for senior skills on day one. If your team uses AI, decide how a junior will learn the judgement the busywork used to teach.
AI agents don't learn from their mistakes on their own. The model is the same tomorrow as it was today, so a lesson only sticks once it's written into the rules the agent reads before it acts. Review every miss, write each one as a single rule with evidence, and make sure someone actually decides which rules go live.
An AI agent you can't stop isn't safe to run. Interruptibility means you can halt one kind of action, or the whole agent, at any moment, without it pushing back or leaving half-done work. Put the stop outside the prompt, make it graceful, and turn things back on one lane at a time.
When several AI agents share the same accounts, files and queues, they will eventually do the same job twice. The fix isn't a smarter model. It's giving every shared thing exactly one owner, and making each agent check the record before it acts.
An AI agent should ask a human before any action that is public, irreversible or expensive. And the yes has to come from the human, in a channel the agent trusts. An approval passed along by another agent is not an approval.
A production agent rarely fails loudly. It goes quiet, or it reports success on work it never did. Give every agent a heartbeat, a receipt for each run, and an alarm for when neither shows up.
Most AI agents still act as you, inside your session. That is why nobody can tell who did what. Give the agent its own identity: named, scoped, short-lived and revocable.
A bigger context window will not save a production agent. The context window is RAM. Memory is what you write on purpose — with a receipt.
Giving an agent a card is easy. Giving it a limit, an expiry, and a list of things it can never do without you is the actual design work.
The useful question isn't whether humans get replaced. It's which layer of the job already has — and what the remaining human is actually for.
An outbound voice agent that actually calls your leads, qualifies them, books meetings, and writes everything back to your CRM — built in a weekend with Vapi, ElevenLabs, and a small amount of glue code.
Why "agents that work on a Tuesday" is the only useful metric in this whole space — and the unsexy plumbing that gets you there.
An open-source scaffold for shipping a real Browserbase + Claude Code agent on a cron, in under 4 hours. Here's how it works, and why it's the smallest useful artifact in agent-land.
Why MCPs and SKILL.md will eat the productivity tool market — and how to be on the right side of that shift.
Most agent evals tell you what you want to hear. Here's how to design ones that actually catch regressions before your users do.
The most valuable thing you can build right now is the thing the market is going to need in exactly a year. Here's how to find that window.
Agents · Engineering · Talks · Podcasts · Interviews · Automations.