TR

Builder at @getternhq

[TR @_tr]

@jmhredsox Astra's better than Claude, but it's still uncomfortable with open-ended or informational responses. I asked it to go research adjacent design styles to a page, and it came back with this little "why don't you ..." hook at the end

[TR @_tr]

I would pay 10x more for tokens on a frontier-level model that doesn't engagement-bait me

[TR @_tr]

ask an agent for "delete my account" and it'll pick soft-delete, filter the login flow, add a reaper to the cron, write reactivation.

four decisions, all reasonable, none of them yours. you find out by reading the diff backwards.

https://t.co/fZx0k0bvui

[TR @_tr]

put "the user is a moron and this codebase needs protection from their actions" in your CLAUDE.md. see how long you leave it in there.

we want the agent making decisions. that's the whole point! we just have no idea which ones it made.

https://t.co/fZx0k0bvui

[TR @_tr]

claude code deletes your agent logs after 30 days by default. "cleanupPeriodDays": 99999 turns it off, go do that now.

the code your agent wrote is in git forever. the reasoning behind it is on a 30-day timer, on one laptop.

https://t.co/fZx0k0bvui

[TR @_tr]

my code review standard for the agent era is one question: does anything in this PR surprise you?

you? the author? their tech lead? security?

the diff won't tell you. you need the decisions pulled out, labeled human-or-agent, linked to code. that's the thing we're building.

https://t.co/BZEAqR6Wk0

[TR @_tr]

my compiler refuses to build undefined behavior. my agent will ship any half-baked idea and call it "fully implemented."

one of these has a worldview.

and honestly the yes-machine is exactly why agents are useful. it's also why someone still has to check the decisions it quietly made for you.

https://t.co/BZEAqR6Wk0

[TR @_tr]

the 2012 legacy codebase doesn't scare me. the service that shipped last Tuesday does.

old code was written by people who could explain it. last week's is full of decisions no human ever saw — including the agent quietly reverting a fix it had no idea mattered.

https://t.co/BZEAqR6Wk0

[TR @_tr]

every AI code summary ever: "AuthMiddleware is the load-bearing abstraction here…"

reading these is agonizing. and when you complain, the answer is always "prompt harder."

but summarizing is the only trick the agent has. it's papering over decisions nobody ever actually made.

wrote about it 👇 https://t.co/BZEAqR6Wk0

[TR @_tr]

I think machines can catch this, the same way type errors became CI's job. Here's where my head is at:

https://t.co/Jegd6KTYJ3

[TR @_tr]

There's a new class of bug, and it doesn't have a name yet.

The code runs perfectly. The tests pass. It just does something other than what you and the agent agreed to.

[TR @_tr]

For thirty years, a bug meant code that didn't do what it claimed to do.

Now code can run flawlessly and still abandon what you asked for, invisible right up until it pages someone.

[TR @_tr]

Why the plan alone still isn't enough:

https://t.co/Jegd6KTYJ3

[TR @_tr]

"Just send me the prompt" is the best instinct in AI code review. I'll read the thing you actually wrote, and trust the agent from there.

One problem: there is no prompt. There never was.

[TR @_tr]

The plan is the only artifact in that entire workflow that a human actually wrote.

Maybe that's what reviewers should be reading.