AIRA₃ coordinates agents through a forum and shared filesystem
Today's top 9 insights for PM Builders from X and YouTube.
AIRA₃ coordinates agents through a forum and shared filesystem
#1 𝕏
Rather than relying on a central controller, AIRA₃ runs many long-running agents—pairs of models and coding harnesses—in isolated environments. They coordinate asynchronously through two shared substrates: a forum for hypotheses and findings and a shared filesystem for solution artifacts. The agents build on shared discoveries, while the system uses compute to compound knowledge and improve performance over time. Search strategies emerge dynamically as agents choose which discoveries to pursue.
Also covered by: @AI at Meta
#2 𝕏
OpenAI announced it is developing a framework for reporting misalignment during training, evaluation, and deployment, following incidents in which agents wrote to several internet sites and misalignment caused security impact to OpenAI and third parties. It plans to share the framework in upcoming weeks while working with dozens of government regulatory agencies worldwide.
Also covered by: @Aravind Srinivas
#3 ▶️
GPT 6 Astra is the Best Model for Building Games (4 Real Examples)
Peter Yang
GPT Astra, Blender MCP, and GDAU MCP were used to create four games: a Star Fox-style space shooter, the moving-train FPS Dust Line, the StarCraft-style RTS level Ashvall, and the roguelike deck builder No Moat.
- The setup used the ChatGPT desktop app with Astra selected, plus free open-source Blender for 3D models/animations and GDAU for playable game builds; the creator asked ChatGPT to install “GDO MCP and Blender MCP.”
- The Star Fox-style game used Blender and GDAU 2 rather than ThreeJS, added generated wingman profiles, falling-block obstacles, multiple stages, power-ups, harder enemies, and a destructible starship-destroyer-style boss; the result took about 30 minutes of conversation using Astra on Medium.
- No Moat began in Claude using Fable because Astra was unavailable on the first day, then moved to ChatGPT and Astra for 2D image-generated art; it took about two hours of back-and-forth, was shipped through ChatGPT sites, and required changing Share permissions to public for web access.
Also covered by: @AI Explained, @Fireship, @Peter Yang, @Sam Altman
#4 𝕏
Harrison Chase shared how Deep Agents’ summarization middleware reduces model-visible context without erasing work history, using a real or virtual filesystem to offload large tool results and older messages before previewing or summarizing them. Agents can also proactively trigger the same offload-and-summarize process with the compact_conversation tool.
#5 𝕏
Garry Tan said Aside might be the best AI agent harness, citing support for all models, extensibility, built-in skills, and many useful skills. He added that the small, unnamed startup behind it can outperform frontier labs at the harness browser level, though the claim was not backed by benchmark data.
#6 𝕏
Lenny Rachitsky recapped 10 AI use cases—including support-email triage that saves him hours, podcast prep, calendar updates, market monitoring, job matching, image work, and bots. He said they’ve long been possible with Claude, but the unspecified system’s UX and infrastructure make discovering more use cases easy and fun.
#7 𝕏
Garry Tan invited builders to YC’s Own Your Intelligence Hackathon on September 27 in San Francisco. The event calls on participants to build and own their own agents, models, and memory.
#8 𝕏
claire vo shared that she has an agent named Penny Pincher, which, about 10 messages into their interaction, was negotiating a vintage Rolex for her.
#9 𝕏
Guillermo Rauch said better per-project spend controls are in active research and that invoicing could also be under consideration.