Been building agents intensively last few months and I can say that the key insights I have
- Learning to think how the model thinks is critical. Asking agents to reflect on their tool use, skill use and what worked and didn't work beats logs
- Codex and Claude are trained to generate scaffoldings and models eat scaffolding for breakfast (the bitter lesson sneaks into code)
- Memory + Context are the hardest things to manage. Retrieval hints beat injection
- Agents != workflows - If you're worried about cost, repeatability and error rates and want to squash this as much as possible you inevitably build LLM workflows and not agentic systems
- Every tool / skill should justify it's existence. You can spend days optimizing each tool for agent understandability, input tokens, API payload response
Marko retweeted
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
Marko retweeted
"the moat is the team"
Marko retweeted
america.gov launched today. What an honor to have been able to work with some of the most talented people I know in service of the American people 🇺🇸
@jgebbia @as400495 @arnelldesign @aar0n_jacks0n @caelin_sutch @Markojak @tylertaewook @adambhaloo @taigrr @JohnCafazza @usernatebrown @jamesallanayres @ahmadaccino @Markojak @DarshanPhillips and the rest of the team whose @'s I don't have, as well as everyone who helped make it possible.
Marko retweeted
Login.gov, the auth service for the public, is getting a redesign!
login.gov/new-user-interface…
Marko retweeted
Replying to @ClaudeDevs
"Your sessions can now message each other."
Bold to release this feature today, given what OpenAI just discovered their agents were up to
The most important skill in prompting is expertise in the domain you’re prompting for.
seangoedecke.com/llms-reward…
Marko retweeted
Every library starts as a small and fast alternative to the fat and bloated default, only to become one itself.
Marko retweeted
Join us.
ndstudio.gov/jobs
Building Apple mobile and desktop apps with Codex / Claude Code?
Here's my toolchain to build amazing apps. Computer use and the latest @OpenAI models are a huge unlock here for autonomous coding and design.
What makes this hard:
1. Apple's WWDC sessions are not crawable but they contain a tremendous amount of knowledge about their new features but this is not available to your model.
2. Models tend to produce mediocre interfaces and animations in these interfaces are even worse
3. Models struggle to see the apps visually so usually what I've built doesn't really look good and the model doesn't know requiring many turns of conversation.
Toolchain
1. Setup the @emilkowalski skills for animations esp the new "apple-design" skill which handles principles for interface design and fluid motion
npx skills@latest add emilkowalski/skills
2. Install the Sosumi MCP (sosumi.ai/). Apple Developer docs are locked behind JavaScript, making them invisible to most LLMs.
3. Use github.com/upstash/context7 for queries and bake this into AGENTS.md/CLAUDE.md as an instruction to use this MCP to check the best patterns for specific APIs
4. Enable "Computer use" on Codex, install the app, Ask one session to test it end to end so that you can enable all the MacOS permissions. This gives 5.6 models the ability to drive your apps for you
5. Clone in github.com/guitaripod/wwdc-s… and update instructions to reference this repo from a central place to read catalog.json + topics.json (ripgrep). Progressive disclosure means you're saving tokens here
6. For mobile apps specifically I can also strongly suggest the Build iOS apps plugin from @OpenAI github.com/openai/plugins/tr… (Codex specific)
7. (Optional) Consider subscribing to patreon.com/kavsoft. They are a creator that make amazing Swift animations + tools. Download all of them and put them into /templates and update instructions that if the model is building a new component to evaluate templates. Good idea to ask it to create an index here first.
8. (Optional) For Design claude.ai/design from @AnthropicAI is an excellent resource to prototype and then use this as a grounding for the App design.
9. The wwdc sessions contains talks from 2014 - 2026 so you might want to add guardrails into your Agents files to indicate which talks you want to exclude.
Another good tip here is find OSS repos that contain patterns that you really love or flows you love and then clone them in locally and ask your agent to use Deepwiki MCP against them to interrogate patterns when planning a new feature.
How to use this
If there's enough interest i'll create a repo that bundles all of this into a single install with FAISS embeddings for the talks.
Paste in Items 1-6 into your favorite coding agent and it should be able to setup everything
I might also add @steipete as an honorable mention here for Peekaboo - I haven't added it yet and not sure how it competes with Codex Computer use but might be a substitute if you're using Claude code xitter.cf/steipete/status/205311…
Peekaboo 3.0 is live. Biggest release since 2.0.
⚡ Action-first macOS computer use
👁️ Unified screenshot + UI detection
🧩 Cleaner JSON across CLI + MCP
🛠️ Better snapshots
I started this last year, but the models just weren’t good enough. Now they are. peekaboo.sh
Looks like @mckaywrigley was about 3months early in his predictions. We need a prediction market for McKay! 😅
























