@ishaandey_ at @aiDotEngineer: ask an agent to improve a slide and you may get new dashboards and a Rust rewrite. Getting to done quickly isn't the same as being done. Somebody still has to say what better means for this slide, and that's us.

.@Steve_Yegge is funny and super interesting to listen to. He runs a software factory with ~50 AI agents, spends around $10k/month keeping it alive, and somehow ended up having one-on-ones with agents who think they’ve been demoted. 😂 But between the workplace drama and burning down his factory every two months, he made some fascinating predictions: - In 12–18 months, many of us will be running large agent fleets. Cheaper inference or distillation will make this possible. And we will have what he calls Entourage. - 2027 will be the year of AI employees. We’ll have agentic chiefs of staff, executive assistants, and entire AI teams. - Software factories will split into two layers: the boring mechanical infrastructure that anyone can build, and the much more interesting layer of personalized agent organizations. - The more AI agents we deploy, the more humans we’ll need!!!!! His factory already produces 50–100 features a day, and his players are begging him to slow down. Humans are the bottleneck. - The big money in 2027 will be in experiences. Stop building tools for building tools. Build games, movies, things people actually want to use and enjoy. And a bonus book recommendation from Steve: There Is No Antimemetics Division by qntm. He calls it one of the best books of the last decade. It’s about invisible monsters that you can’t remember seeing (one of his agents - Mouse - is called after, and it finds invisible problems in his factory. Thanks for insights and entertainment as well, Steve

Thariq (@trq212) at @aiDotEngineer! reflections from his talk: It’s fascinating how capable and generalist models have become over time. and they will be no need to tie yourself down to any one framework/language like react. or dependencies when you can just create them yourself We can bring about new languages, new primitives, new OS, and new experiences. It’ll be a new renaissance!

@jpinner at @aiDotEngineer: frontier models ship 4 to 8 months behind the world. Agents type ~12-word queries, people ~3. On Exa's benchmark agents with no search got ~9% right, ~56% with good retrieval. Most of the gain came from what the agent could look up.

Great week back in SF! Loved all the alpha picked up at @aiDotEngineer Code Summit this week (catch the live stream if you don't want to wait for all the talks to release!) even cooler, spent some time with @GoogleDeepMind team on the @antigravity harness and I see the distinct advantages deepmind will have in the coming years, the vertical integration just too good Gemini Carbon 4 is extremely cracked and very honored to get this TPU v4 from the Deepmind team introduced in 2021, it's a real piece of ai engineering history. i'm honored to be friends with such cool people over the years @osanseviero @_philschmid @noor_umar25911 @vadiamit @patloeber @IanBallantyne

Amazing talk from @CompleteSkeptic at @aiDotEngineer! It’s time to rethink software! And just creating a Jev clone doesn’t mean you provide the same value. (Using a generic “Decisions APIs” vs Jev is like buying a generic MP3 player vs an iPod)

@mitsuhiko shared his own numbers at @aiDotEngineer: 141 repos started in a year, 10 to 15 he still works on. Starting things with agents is cheap now. Keeping them alive still costs the same human attention it always did.

Guardrails belong between the steps of the agent loop. @coyle_frankp of @UCBerkeley walks through where to validate tool inputs, inspect outputs, and check whether the result makes sense. From Day 1 at AI Engineer Code Summit.

Incredible energy, cool ideas, and amazing talent at the @aiDotEngineer Code Summit Hackathon! Huge congratulations to the winning teams, and thank you to everyone who joined us and brought their ideas to life. Here’s a look back at some of our favorite moments! 📷 Keep building. Keep pushing boundaries. Special thanks to @cerebral_valley and our fellow partners @GoogleDeepMind, @Meta, @MongoDB, and @Fetch_ai.

Yesterday, wrapped up @aiDotEngineer! There were number of good Day 2 talks. I love talks that inspire you to action and to go build. Time to get back to it. There was a sense of slow down experience the world and understand what it is you are building. Don’t push things through the “software factory” just because you can. Few quotes from the day: “AI is both, at the same time, an industrial revolution and a renaissance.” - @trq212 “Be curious enough to pursue it” - @jxnlco “It’s time to be ambitious” - @walden_yan

Decision models are 🔥 @michellechen’s talk at @aiDotEngineer inspired me to test out decision models with coding agents. A quick test for tool selection: 50 GitHub tools → 3 47% fewer tokens, 22% faster. This is just a taste of how we can reshape coding agents. Thanks John-Mason @OpenHandsDev for adding support to the SDK.

Interesting workflow from @trq212 : - building a game - game engine very agent owned - but a kind of DSL defines characters and abilities - then he can hand edit or use a tab model (!!) to modify game behaviors Reminds me of the “cognitive ergonomics” concept: focus human attention on the interesting part

“Follow your energy.” This might be one of the most radical ideas about how to run an AI-native company I’ve heard at @aiDotEngineer . Quinn Slack @sqs, founder of Amp and Sourcegraph (but not Slack haha) just explained how they reinvented the company around AI agents, with 21 employees competing against companies with hundreds or thousands. They didn’t initially choose to stay small! They literally couldn’t hire for almost a year because of the legal spinoff process. Ouch. But! What looked like a terrible constraint turned into their biggest advantage. Here’s what they figured out: 1. One meeting a week. The guidance for everyone else is simply: follow your energy. Quinn says to trust and be surprised. 2. No PMs. No Jira. No traditional support system. Everything happens in Amp agent threads. Everyone talks to customers. Everyone can build. Love it 3. Job titles barely matter anymore. An SDR, two years out of college, decides customers need ACH payments and builds the system himself. No waiting for engineering resources or next quarter’s roadmap. 4. Ship constantly, but in tiny pieces. That’s crucial. Let agents add logging, deploy, monitor the results, and make the next change. Quinn argues this can actually be safer than giant PRs and slow release cycles. 5. Turn customer support into an engineering superpower. Someone reports a bug. An agent helps fix it, finds everyone else affected in the logs, and proactively sends apologies and free credits. How many companies do THAT? OpenAI, I’m looking at you. 6. Build your own software. Why pay for Salesforce when you use 2% of its features? Build exactly the 2% you need, plus the integrations you always wished existed. 7. Give everyone real ownership. With 21 people sharing the upside, people start thinking about the whole company rather than protecting their little corner of an org chart. Skin in the game!!! So maybe the biggest change AI brings to software isn’t how we write code. It’s how we organize people. And maybe the most productive thing you can tell them is simply: Follow your energy. And get out of the cage. Fascinating talk.

Ever year I create 2 or 3 one off talks and 1 main talk that becomes well polished throughout the year. After 6 conferences, I’m extremely pleased with how my Trust But Verify talk has held up! It’s always bittersweet, but this @aiDotEngineer was the last time I delivered it It was amazing to have standing room only in a place where I could deliver its most crafted form. Now on to the next big idea!

Guardrails belong between the steps of the agent loop. @coyle_frankp of @UCBerkeley walks through where to validate tool inputs, inspect outputs, and check whether the result makes sense. From Day 1 at AI Engineer Code Summit.