Recommendation · MoneyDaksh Gupta recommends that for autonomous end-to-end testing you will need a dev box, PR branch dependencies where you're running the app, and some type of flow map that says 'this is the auth code, and then this is the signup flow, and so on.'
AI Engineer · How to Build an AI Code Review Agent - Daksh Gupta, Greptile
Recommendation · MoneyDaksh Gupta recommends using a small model for finding context, a lightweight computer-use model for testing user flows, and a frontier model for analyzing bugs, applying the review process, and submission.
AI Engineer · How to Build an AI Code Review Agent - Daksh Gupta, Greptile
Recommendation · MoneyDaksh Gupta recommends using vision models to catch UI issues that code review misses, such as a button that technically appears but is hidden on iPhones or not in the first fold.
AI Engineer · How to Build an AI Code Review Agent - Daksh Gupta, Greptile
Recommendation · MoneyScrape legitimate community Q&A from chat platforms into GitHub Discussions, then mirror that content onto your public website so it becomes indexable and LLM-trainable.
Latent Space · Supabase’s Plan for a Billion AI-Built Apps
Recommendation · MoneyYou cannot trust the manager agent to self-direct reading children sessions; Ryan Carson says you must build a deterministic script or automation for it.
AI Engineer · How One Founder Manages 20 AI Coding Agents at Once - Ryan Carson, Untangle
Recommendation · MoneyFeed all finalist videos into an agent (Devin) and ask it which candidate to pick.
AI Engineer · How One Founder Manages 20 AI Coding Agents at Once - Ryan Carson, Untangle
Recommendation · MoneyWhen reviewing a PR video, read a summary of what the thing is supposed to do, watch the video, and engage with it. This is the part where you want to cut corners, but you have to engage as a product owner and pay attention.
AI Engineer · How One Founder Manages 20 AI Coding Agents at Once - Ryan Carson, Untangle
Recommendation · MoneyTereza advises hosting open models on dedicated compute, noting people sometimes forget open models can do caching too, so you can take the same advantage of caching that LLM labs get.
hedged
AI Engineer · Rise of the Software Factory — Tereza Tížková, Factory
Recommendation · MoneyTereza advises that organizations can give different permissions and different default models to different people — for example, marketing, sales, and engineers can each have different default models — which she calls very useful.
AI Engineer · Rise of the Software Factory — Tereza Tížková, Factory
Recommendation · MoneyTo build an eval from production traces, collect a bunch of production traces from your agent and get a human to annotate just about 50 of those trajectories with very detailed feedback.
AI Engineer · Beating RL With Reflection: GEPA and Optimize Anything — Lakshya A. Agrawal, GEPA
Recommendation · MoneyLakshya A. Agrawal advises: use single-problem mode for one problem (e.g., a single matmul kernel), multitask search mode for n related problems where information might transfer (e.g., matmul plus dot product kernel), and build-a-skill mode when optimizing on a set number of problems but deployment will face many new problems, since it cares about generalization.
hedged
AI Engineer · Beating RL With Reflection: GEPA and Optimize Anything — Lakshya A. Agrawal, GEPA
Recommendation · MoneyRoman Ugarte describes collecting a rich set of tasks people actually give the bot, quantifying them, and tracking week over week how they hill climb on the most important dimensions behind the scenes. When a workflow fails - for example the mouse lacks fine enough control to click part of a Salesforce dashboard - the concrete case goes back to the core infrastructure team, which he found far more tangible than watching a number on a dashboard creep up.
Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)
Recommendation · MoneyRoman Ugarte recommends reframing product descriptions from the old-school software tendency of 'GrokBot now has' (new buttons and dropdowns) to 'GrokBot can now' (capabilities), which he calls a much more human way of describing them. He also recommends killing as many pixels as possible, letting the bot manipulate things behind the scenes, and never exposing an interface for creating automations - users should just tell the bot in natural language, e.g. 'remind me at 8 a.m. every day'.
Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)
Recommendation · MoneyRoman Ugarte says a very big group with a 6-12 month long vision would not have landed where they ended up; instead the team needed to make a lot of micro decisions every day.
hedged
Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)
Recommendation · MoneyRoman Ugarte says it was important for the core team to sit on onboarding calls for 20 minutes to directly observe user confusion, and that when they saw a user confused they prioritized solving it by the next day's onboarding.
hedged
Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)
Recommendation · MoneyRoman recommends onboarding early users manually — the team onboarded two to three hundred people, including the host — to learn from their experience and build a better product.
Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)
Recommendation · MoneyFor power users, Roman recommends creating a scaffold where GrokBot artifacts live and having it write to a single store so outputs are easy to organize; he personally runs frequent digests pushed to a database that he reads every day.
Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)
Recommendation · MoneyIrina's team recommends that if source attribution disappears, the policy change to search over might be to preserve source in memory writes and summaries.
hedged
AI Engineer · Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati
Recommendation · MoneyIrina's team recommends scoring the run's behavior by asking: did the information spread as expected, did source attribution survive, did uncertainty stay uncertain, and did agents act on what they actually knew.
AI Engineer · Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati
Recommendation · MoneyIrina's team recommends assigning every memory an importance score. When an event's importance score crosses a threshold, store that memory in a separate cache so important context can be retrieved better later. Irina illustrates this with the analogy that you won't remember what you had for dinner a few days ago, but you definitely remember if someone was murdered.
hedged
AI Engineer · Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati