Everyone is shipping agents. Almost nobody is shipping guardrails.
Welcome to the first one. The plan is simple: one email a week about AI — what I'm using, what I've worked out, and what's genuinely worth your time. If a week is thin, you get a short email rather than a padded one.
The thing I keep noticing
Every tool this month wants to be an agent. Very few of them have an answer for what happens when the agent is confidently wrong at step four of nine.
The setups that have actually held up for me share one property: the agent proposes, and something deterministic disposes. A test suite, a schema, a type checker, a diff you look at. Not another model grading the first model.
If the only thing checking the output is the same kind of thing that produced it, you have not added a check. You have added a second opinion. |
What I'd actually try this week
Take one task you have already automated with a model and add a single non-negotiable gate to it — something that can only pass or fail, with no judgement involved. Watch how much your trust in the whole pipeline changes from that one addition.
One thing to skip
Any tool whose demo is another tool being demoed. If the pitch never reaches a real artifact — a file, a deploy, a passing test — there is usually a reason.
That's it for this week. Reply if you disagree; it comes straight to me.