Building AI that runs itself. Senior Engineer at @TradeRepublic. Speaker. iOS, Swift, Rust and Python. Making https://nitter.cf/t.co/Kq7FisXWls - https://nitter.cf/t.co/IJLnfrldRk
Berlin
Joined April 2009
- Tweets1.2K
- Following596
- Followers766
- Likes985
Limits are per person because accountability is per person, and I suspect most companies keep them until someone decides who answers for the agent when it ships a bad change.
Scope creep from coding agents is worse than them doing too little, since every extra line still needs someone to review it.
I'm pretty sure most agents on production systems run with the token of whoever set them up, so they can do anything that person can. They should get their own account with its own permissions, the same way any other service does.
Running models locally turned RAM into the spec I check first. I wonder if laptops start getting sold by memory size the way they used to be sold by CPU speed.
One bad incident with AI-written code in a big company usually turns into a new rule for every team, another approval step or a tool nobody can use anymore.
In a small team the person who shipped it fixes it the same day and moves on, and I have a feeling that's why small teams keep trusting these tools while big companies slowly fence them in.
The problem space might be infinite, but people's attention isn't, and to me that's where most of the competition happens now.
In an infinite problem space there is no competition. There are so many challenges in the world that you and your competitors can't solve all anyhow. There is enough for everyone. Operate from that and all of a sudden you see a billion opportunities. There maybe is no "moat". However, there is definitively an unlimited space of opportunity and solving humanities challenges that no singular entity can fill.
Blanket retries are how one slow dependency turns into a full outage, and my bet is an agent left on its own would sprinkle them everywhere too.
The menu bar and the Dynamic Island are hard places for an LLM to reach, and my guess is that's where simple utilities keep making money.
A new coding model launch says very little until someone runs it inside a real harness. The same model can look great in one setup and lose the plot in another, and I wouldn't be surprised if the harness is half of the result.
Big companies still can't get budget to migrate off an old Java version, so the rewrite is where I think this stalls.
Quantization is a tradeoff hosted APIs make for us and rarely disclose. Running the model myself puts that call in my hands, and my take is that control is worth more than the price gap.
Markdown versions of websites remind me of the old m. subdomains from the early mobile web. Those got folded back into one responsive site, and I'd expect the same here once agents read the regular page well enough.
Doing a task exactly the same way every time is something an LLM still doesn't offer, and I'm pretty sure that's where a lot of simple apps keep their paying users.
Letting an agent read my texts so it can get past 2FA feels like it defeats the whole point of having 2FA.
When AI writes more of the code, a developer in a small team ends up making the product and design calls too, because there's nobody else around to make them. I have a feeling that's the skill small team developers will get good at first.
Throwaway apps make sense for one person and one task. In a company my bet is Figma and Linear stay, because their whole value is everyone looking at the same file and the same board.
The single data store part I agree with.
I used to think personal software built with AI was the future. Everyone builds their own Figma, Linear, video editor, etc.
I think that’s wrong now. Short lived purpose built software for the task at hand seems likelier with a single data store to keep source of truth/outcomes/output. Especially as models get better and faster.
If you’re a builder, build for interoperability.
Getting bought before the labs ship the same feature is probably the best ending for a tool built on top of coding agents.
Conductor is joining Meta!
It's been so cool to see their evolution from pivot hell during the S24 batch, to building a product that they themselves wanted to use (back in the stone age of 2025 when people still coded by hand!), to building a community and movement around their product.
All of their success is so well deserved, couldn't happen to better people.
Congrats @charlieholtz @jacksondecampos and the entire @conductor_build team!
Circling a spot and drawing an arrow is how designers have always given feedback, so I wouldn't be surprised if markup editing ends up beating any written prompt for image edits.
Flat subscriptions win for anyone coding all day, at least while the labs keep subsidizing their heaviest users.