How to build an agentic loop with Claude Code

#1 𝕏

Santiago demonstrates how to build your first agentic loop with Claude Code by running a single CLI command—`claude -p "Write fibonacci(n) in a Python file…" --allowedTools "Read,Write,Edit,Bash(pytest…)" --max-turns 15`—to auto-generate and pytest-verify a Python Fibonacci i...

#2 𝕏

LlamaIndex 🦙 built a template for Vercel’s new Eve agent framework that pairs read-only filesystem tools (path resolution, directory listing, file reading) with LiteParse to output clean, structured Markdown.

Also covered by: @Guillermo Rauch

#3 𝕏

Guillermo Rauch: Sandbox can now run Docker and FUSE unconstrained on instant-boot microVMs as the foundation for Fluid Compute, and he shipped a 10-line S3-backed filesystem for agents.

Also covered by: @Guillermo Rauch

#4 ▶️

I Tested Gemini Spark: What Google’s AI Agent Can Actually Do in 21 Minutes

Peter Yang

#5 in

Peter Yang shares a three-step workflow to maximize Fable before July 7—prep with cheaper LLMs, plan in Fable and execute with another model, then assign medium-effort tasks with a bit of oversight—and links to a tutorial on five practical Fable use cases.

#6 in

Omon Eni spotlights Carl Vellotti’s free, five-module course that turns Anthropic’s Claude Code into a hands-on PM operating system. PMs clone a repo, open their terminal, and in three steps learn by doing—writing PRDs, running data analysis, and building strategy docs.

#7 📝 Simon Willison

Open Source AI Gap Map - Current AI launched an Open Source AI Gap Map indexing hundreds of open-source AI artifacts and released the underlying dataset on GitHub under an MIT license. Simon highlights the dataset (1,184 YAML files) and suggests exploring the tracked GitHub repos with Datasette Lite.

#8 📝 Ampcode Chronicle

More Orb Sizes - Amp now offers four orb sizes: a0.tiny (1 CPU, 2GB memory, 40GB disk, $0.10/hour), a0.small (2 CPUs, 4GB, 40GB, $0.21/hour), a0.medium (8 CPUs, 16GB, 40GB, $0.83/hour) and a0.large (16 CPUs, 32GB, 40GB, $1.66/hour) which is the default; orb storage has been doubled from 20GB to 40GB at no extra cost and you can change a project's orb size in Project Settings.

#9 in

Marc Baselga highlights OpenAI’s update that makes ChatGPT better at health queries—from sleep advice to symptom analysis and medication guidance—for 230 million weekly users. He argues the real innovation lies in the rigorous safety and reliability work behind the scenes.

#10 𝕏

Alexandr Wang previews the next Muse Spark update with major coding and agentic capability improvements, rolling out soon to Meta AI and the new API.

#11 𝕏

Yann LeCun warns that the greatest AI risk is concentrating power in a handful of companies or states—creating medieval-style control over information and economic tools akin to the Ottoman ban on printing presses.

#12 in

Udi Menkes reports from AIE that 94% of teams build closed AI models but struggle with evaluation (still relying on “vibe reviews”), and DSPy’s founders push for formal I/O contracts over ad-hoc prompts to make systems maintainable.

#13 𝕏

Harrison Chase launched OpenWiki, which hit 1.7k GitHub stars in two days, and the top request is to broaden it beyond coding to integrate sources like Notion, Gmail, Slack, GDrive, internet search, and more.

#14 𝕏

Logan Kilpatrick notes we currently only have Omni Flash available and warns that performance insights won’t necessarily translate unless your usage distribution matches theirs.

#15 𝕏

Logan Kilpatrick clarifies the leaderboard is run by a third party and he only saw the results when they were published. He urges skepticism and deeper evaluation of the data.

#16 𝕏

Santiago cautions that unit tests alone only suffice for simple demos and advocates using multiple validation methods for more complex scenarios. He argues that writing and validating code tap into different skills, so it’s fine to use the same model for both.

#17 𝕏

Thariq explains how he surfaces his own “unknowns” through HTML artifact analysis to craft sharper prompts for Fable, walking through a step-by-step process with live examples.

#18 in

Greg Isenberg warns Claude’s Fable 5 is free only through July 7 before switching to pay-per-use at $10/$50 per million tokens and already burns tokens twice as fast with a 50% weekly cap.

#19 📝 Simon Willison

Judgement - Simon describes a tip from a Fireside Chat to let coding agents like Claude Fable use their own judgment (e.g., when to write tests or to delegate tasks to lower-power models). He demonstrates prompting Claude Code to save a memory instructing subagent delegation and shows the stored memory and recommended application pattern.