Explore

Set up a self-improvement loop where agents analyze failure modes and propose PRs to the factory definition

Zach Lloyd recommends setting up a self-improvement loop where agents look at failure modes, aggregate them, look for patterns, and come up with ideas to improve the underlying factory definition — closing the loop by having the agent propose a PR to the factory definition when it identifies a recurring failure mode. He illustrates the agent's framing as 'I saw this fail 37 times' and 'we can actually fix this with a PR to the factory definition'.

AI Engineer · How to Build a Software Factory That Improves Itself - Zach Lloyd, Warp

Create your own benchmarks from real coding agent tasks

Zach Lloyd recommends creating your own benchmarks by taking all of your real coding agent tasks and trying them against different models, measuring how it goes. As a starting approach, he recommends building a benchmark set by hand, picking a set of small, medium, and hard representative tasks. Alternatively, you can ask the factory itself to build the benchmark set, since it is aware of all past sessions and everything is stored in code — the factory can look at prior sessions and identify interesting ones. You can also take any past run from your system and add it to a benchmark suite, and curate a benchmark of representative tasks from your own factory to measure model performance on your own data. He gives the example of building a suite of front-end tasks to answer which model isbest

AI Engineer · How to Build a Software Factory That Improves Itself - Zach Lloyd, Warp

Set up online evals (scores) with LLM-as-a-judge

Zach Lloyd recommends setting up scores — online evals — where a set of agents looks at all the runs going through your system and uses LLM-as-a-judge to figure out whether the runs are going well. At Warp, they have a set of agents looking at all of these runs, and a scoring mechanism that evaluates them using LLM as a judge.

AI Engineer · How to Build a Software Factory That Improves Itself - Zach Lloyd, Warp

Frontier models for critical tasks, smaller models for the long tail of tokens

Daksh Gupta recommends using frontier models for the most critical tasks (analyzing bugs, applying the review process, submission) and smaller models for the long tail of tokens (exploratory, file reads).

AI Engineer · How to Build an AI Code Review Agent - Daksh Gupta, Greptile

Give agents access to JIRA, Linear, and docs stores

Daksh Gupta recommends giving the agent access to JIRA, Linear, and a docs store like Confluence or Notion, claiming you can get remarkably far with just these two types of context: docs and tickets.

AI Engineer · How to Build an AI Code Review Agent - Daksh Gupta, Greptile

Use model inversion in your code review process

Daksh Gupta highly recommends using some type of model inversion in your code review process: if you used Claude Code to write the code, use a GPT model to review it, and vice versa.

AI Engineer · How to Build an AI Code Review Agent - Daksh Gupta, Greptile

Use Multigress for Postgres sharding without changing queries

Multigress was ground-up rewritten so you don't have to change any queries when you shard; it was stress-tested with 1,000 sequential failovers and clobbering tests with zero data loss before release.

Latent Space · Supabase’s Plan for a Billion AI-Built Apps

Run long-running serverless compute co-located with your database

Supabase Compute lets you run long-running serverless workloads (chunking, embedding, agent jobs) with direct Postgres connections, and it is not limited to Deno — Rust, Python, or full containers can run inside it.

Latent Space · Supabase’s Plan for a Billion AI-Built Apps

Use sharding to scale OLTP Postgres into the petabyte range; treat analytics separately

Use sharding to scale OLTP Postgres into the petabyte range rather than trying to make Postgres an analytical database. Treat analytics workloads separately from transactional (OLTP) workloads — use the right technology for the right job. Splitting transactions across two servers forces fetching from object storage, which will eventually slow down your database. You can still store data directly on object storage and read from it via a different route, just don't fuse OLTP and analytics into one system.

Latent Space · Supabase’s Plan for a Billion AI-Built Apps

Start prototypes on Supabase Spark and upgrade to full Postgres when the project takes off

Start prototypes on Supabase Spark (SQLite-compatible small instance) and upgrade to full Postgres when the project takes off — one database for any idea. Zero-downtime migrations between backends are a need-to-have, not a nice-to-have, because apps can spike from zero to front-page-of-Hacker-News traffic overnight.

Latent Space · Supabase’s Plan for a Billion AI-Built Apps

Move to code-first, agent-native: make database schema fully declarative in files

Supabase's big focus this year is moving to code-first, agent-native: launching a Supabase instance on your local machine gives you a folder representing your database. Use the Supabase CLI to get a virtual file system representation of your database — infrastructure as code. Make database schema fully declarative in files (tables, policies, comments) so agents can iterate fast on the file system instead of reaching out to a running database. Supabase is releasing pg-donta, a declarative schema library usable with any Postgres database.

Latent Space · Supabase’s Plan for a Billion AI-Built Apps

How to perform the incline drag curl

Set the incline bench to 60 degrees, use light weights, start with elbows back behind the body, curl by bending only at the elbow while twisting the dumbbell into supination, allow the elbows to come forward slightly with an extra twist at the very end, then come down slow and re-establish elbow positioning, contracting the triceps at the bottom for a deeper biceps stretch.

Method, not a verified protocol.

ATHLEAN-X™ · The ONLY 2 DB Biceps Exercises You Need (NO, SERIOUSLY!)

How to perform the preacher curl on the 60-degree bench

Perform the preacher curl on the same 60-degree bench angle, press the entire back of the triceps against the pad, push back into the pad to engage the lats and support the shoulder girdle, keep tension in the bottom position, and twist the wrist into supination as much as possible at the top.

hedged

Method, not a verified protocol.

ATHLEAN-X™ · The ONLY 2 DB Biceps Exercises You Need (NO, SERIOUSLY!)

Save money on AI spend with caching, no spending limits, and smart routing

Tereza recommends three levers for saving money on AI: use caching to stop prefilling context every time, set no limits on spending but require people to show results when they spend a lot, and use smart routing between LLMs to optimize tokens instead of spending so many tokens.

AI Engineer · Rise of the Software Factory — Tereza Tížková, Factory

Onboard early users manually and fix the worst pain point before the next session

Roman Ugarte describes manually onboarding 200-300 early users, sitting on a 20-minute call watching for confusion or technical issues, and immediately fixing the worst pain point so the next onboarding goes better.

Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)

Build new products from scratch with a small, isolated team

Roman Ugarte recommends starting completely from scratch rather than adding features to an existing product, and creating a small team that goes off into a cave for about a month with a single objective, completely isolated from the rest of the company.

hedged

Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)

Onboard early users manually and fix pain points within 24 hours

Roman Ugarte describes a two-week period of intensive manual onboarding where the team onboarded 200-300 early users via calls. Whenever they observed an onboarding pain point, they fixed it immediately - within 24 hours - because the next person would be onboarded the following day.

Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)

Give the agent hardware context, since low-level GPU features change generation to generation

Tejas Bhakta says you really need to make sure your agent is aware of the hardware — e.g., on a B200, the warps, Tmem, TMA, and so on. These low-level operators change generation to generation (H200 won't have Tmem, which is new with B200), which is why this context is needed. In practice this is a bunch of MD files you give the agent so it has context.

AI Engineer · Autoresearch Made Our Models 3x Faster — Tejas Bhakta, Morph

As the human, review profiler output and flag dumb high-level approaches for the agent to fix

Tejas Bhakta advises that the human's job when reading a profiler is to look at the top and identify dumb approaches — for example, recognizing that loading 32K chunks into context is unnecessary for DeepSeek attention and should only be done every 32K instead. At a high level, all you have to tell autoresearch is that the top method is dumb and to pipeline it instead; everything else like sizing, chunk sizing, and context chunks should be decided by autoresearch.

AI Engineer · Autoresearch Made Our Models 3x Faster — Tejas Bhakta, Morph

Give the agent model context, or it will hallucinate the attention mechanism and produce useless kernels

Tejas Bhakta says you also need to make sure your agent has context of the model — every new model like DeepSeek Flash comes out with new tricks, such as DeepSeek v4's two new attentions: compressed sparse attention and hierarchical compressed attention. If you don't give the agent model context, he says the model will 100% hallucinate the actual attention mechanism and you will get useless kernels.

AI Engineer · Autoresearch Made Our Models 3x Faster — Tejas Bhakta, Morph