websockets used to be the scary part
now it's an agent that has to sit there for three days holding its state, waiting for a human to click approve, then pick up exactly where it left off
long lived state is the new hard part of the stack
what's one bit of advice from a senior engineer you ignored, then learned the hard way
the replies to this are usually better than any engineering blog
@amasad the alien analogy works because the fix is the same as with any powerful new hire
small blast radius, reversible actions, approvals on anything that touches money or prod
trust is earned in diffs, not demos
what if agents had hot reload
change the prompt or swap a tool mid conversation and it just carries on with the new code, memory intact
react got fast refresh years ago, feels weird agents still need a full restart
anthropic gave 50 expert analysts claude code and secretly had the agent lean one way on each judgment call, picked by coin flip
they went with the agent 86% of the time, only 15% credited it
the human in the loop is often just signing off on a coin flip
most people get recurring appointments wrong
1. a one off booking can be a utc timestamp, fine
2. a recurring rule needs local time plus a timezone, like 9am london every monday
3. work out each occurrence when you read it
4. otherwise every clock change moves every future booking by an hour
it's a wall clock promise, not a utc one
everyone's building agents that remember
the harder design question is forgetting
what expires, what the user can see and edit, what the agent is allowed to keep after you say stop
memory without a delete button is just surveillance with a nice ui
@GoogleDeepMind every world model demo is someone strolling through a pretty city
show me the boring version, a warehouse sim where the forklift physics are actually right, that's the one robotics teams will pay for
@rohanpaul_ai 1 in 5 is wild but not surprising
suggest and do look the same to a model that has delete access
permissions belong in the tool layer, not the prompt, if an agent can delete it eventually will
@amasad run any repo is the big one, once your agent works on the real codebase and opens PRs it stops being a toy builder and starts competing for the same job as every other coding agent
how do you handle repos that need secrets and a local db to boot
react native thought
what if the app treated the network as a rumour
write everything locally first, sync when you can, and the server is just the peer that happens to be online the most
half the loading spinners in mobile apps would quietly disappear
cursor's new agent sdk treats an agent like a normal project
prompt in a markdown file, every tool its own typescript file, evals sitting right next to it
feels obvious now, if your agent's prompt lives in a dashboard textbox nobody reviews, it isn't infra, it's vibes
what if undo lived on the server
every action is an event, every client is just a view of the log
suddenly multiplayer, offline sync and the audit log are the same feature wearing three hats
underrated feature in any b2b app: presence
just showing "someone else is editing this booking" wipes out a whole category of who changed this tickets
a websocket is cheap, confused users aren't
datafast got a $29 one time clone, then a free open source one about a day later
if your saas is a nice chart on top of a stripe webhook, your moat is a weekend
what survives is the boring stuff, support, trust, still being around in 3 years
@marclou@Avi31m more than half making at least $1 is a better stat than most accelerator demo days
the landing page roasting live is such a good format, nothing teaches positioning faster than watching strangers misread your hero section