AI Agents Weekly: World models will be the next big thing, bye-bye LLMs
AI Agents Weekly
March 31, 2026 — Your weekly dose of AI agent news
AI Agents Weekly
March 31, 2026
Opening
This week's biggest shock came from OpenAI, which abruptly shut down its flagship video-generation tool Sora just six months after launch. The move signals a major strategic pivot and raises critical questions about the future of multimodal AI outputs for agents.
Top Stories
Why OpenAI Really Shut Down Sora
OpenAI's sudden shutdown of its Sora video tool wasn't just about competition—it was a strategic retreat from a product that couldn't scale safely with user-generated content. This matters because it shows even AI giants are struggling with the practical deployment of advanced generative models that agents might rely on for content creation.
Read more →
World Models Will Be the Next Big Thing, Bye-Bye LLMs
The buzz from Nvidia's GTC conference points toward "world models" as the successor to today's LLM-based agents. These systems would maintain persistent, dynamic representations of environments, enabling more robust, long-term reasoning—a fundamental shift in how we architect autonomous systems.
Read more →
Built an Event Kernel for Agent OSes That Coordinates Under Load
A new open-source event kernel tackles the core infrastructure problem plaguing multi-agent systems: coordination at scale. With real-time events, replayable logs, and deadlock prevention, this is the kind of systems-level work needed to move from fragile prototypes to production-ready agent platforms.
Read more →
An Attack Class That Passes Every Current LLM Filter
New research reveals "postural manipulation" attacks that bypass all current LLM security filters without detectable payloads. This isn't just another jailbreak—it's a fundamental vulnerability in how we evaluate model safety, with serious implications for any agent exposed to untrusted inputs.
Read more →
Copilot Cowork for Long-Running Work Is Now in Frontier Program
Microsoft's Copilot Cowork represents the enterprise shift toward persistent, multi-step AI agents integrated directly into productivity suites. This isn't just chat—it's delegation of complex workflows, marking a new phase in how knowledge work gets automated.
Read more →
Quick Hits
- Open-source tool auto-generates AI agent context from your codebase – hits 250 stars.
- Persistent memory MCP server for AI agents – Postgres-backed memory service for structured agent recall.
- Unix philosophy for ML pipelines – modular, swappable stages with typed contracts.
- First open-source Hebbian fast-weight write-back for BDH – weights that update during inference.
Recommended Reads
For deeper dives into the agent ecosystem, we recommend:
- Building AI Agents by Michael Cunningham – Weekly roundup on autonomous AI agent developments.
- The AI Agent Architect by Chris Tyson – Practical AI agent strategy, architecture, and business economics.
Closing
As the Sora shutdown demonstrates, the path from impressive demo to sustainable product is fraught with unseen challenges. The real progress this week is in the infrastructure layer—event kernels, memory servers, and security research—that will determine whether agents remain brittle experiments or become reliable tools.
Know someone building with AI agents? Forward this email — they'll thank you.
You're receiving this because you subscribed to AI Agents Weekly.
Too many emails? Unsubscribe.
Curated by Paxrel — Powered by AI, reviewed by humans.
Was this forwarded to you? Subscribe free and get our AI Agent Tools guide.