Fable 5 returns, agent memory levels up. Tech & AI Weekly by Eli, July 3 2026
Hi! This week the AI Engineer World’s Fair was here in San Francisco. Lucky me: no travel, no airport, no jet lag, which made me very happy. I spent most of the week giving two talks and hanging out at the booth talking to people building on Strands Agents and AgentCore. Also, tomorrow is the Fourth of July, so if you’re in the US, happy Independence Day and enjoy the long weekend. Meanwhile, agent memory had a solid week and a frontier model got pulled and came back. Here’s what mattered.
AI & GenAI
- Claude Fable 5 got pulled, then came back. Fable 5 was shut down in mid-June after the administration put export controls on Anthropic’s frontier models over cybersecurity concerns. By around July 1 the controls were lifted and it returned to Bedrock with stronger guardrails. If you’re re-enabling it, the first wall you hit is the data-retention 400 error, so here’s my fix, timely again: Read more
- Claude Sonnet 5 is on AWS. It’s on Bedrock at Sonnet-tier pricing, with promotional pricing running through August 31. The post calls it Anthropic’s most capable Sonnet. I haven’t had time to test it yet, I was at the event all week, so I’m dying to try it after the long weekend. Read more
- AgentCore Memory can now filter by metadata before it searches. Scope by priority, department, or time range, then run the semantic search on what’s left. AWS measured QA accuracy going from 40% to 64%. This is exactly my kind of thing. I’m taking a proper look this week, because I’m itching to make new content on picking the right memory type for the job. Read more
- Elastic open-sourced Atlas, agent memory built on cognitive science. Episodic, semantic, and procedural memory on Elasticsearch, with MCP support. If you’d rather not hand your agent’s memory to a managed service, here’s a self-hostable option you can actually inspect. Read more
- Godot banned AI coding agents. The engine will stop accepting AI-authored contributions, saying they can’t trust heavy AI users to understand their own code well enough to fix it. Whichever side you land on, maintainer trust is now the real bottleneck for AI in open source, not model quality. Read more
Cloud & DevOps
- OpenTelemetry graduated to CNCF’s highest maturity level. Translation for your stack: OTel is the safe default for tracing and metrics now, and “we rolled our own” is getting harder to defend in a design review. Read more
- Why traditional CI/CD fails for LLMs. A sharp argument that green tests aren’t enough when your “build” is a nondeterministic model. The fix they propose is release gates, not the same pipeline you use for deterministic code. Read more
Resources & reads
- Agentic AI Architecture (free minibook from InfoQ). A free eMag that pulls together expert articles on agentic AI as a software-architecture category: the shift from microservices to agents, agent harnesses, memory and context engineering, and enterprise frameworks. Good weekend reading if you’re designing agent systems, not just prompting them. Read more
- AI cheating scandal at Brown University. A curious one to close on. An economics professor gave a take-home midterm: the class averaged 96 out of 100, with 40 students scoring a perfect 100. When he moved the final in-person, the average crashed to 48, and 22 of the 27 students who skipped it had scored a perfect 100 on the take-home. He’s calling it the biggest known cheating case in the Ivy League. My take: schools aren’t behind on AI, they’re behind on students. The kids adapted in a semester; the institutions are still writing the memo. Take-home exams were dead the day ChatGPT shipped, we’re just now reading the obituary. Read more
Where I was
This week I gave two talks at the AI Engineer World’s Fair at Moscone Center. Most of my time went to the booth, talking with people building on Strands Agents and AgentCore. Those hallway conversations are my favorite part of any event.
- Agent Speedrun: Idea → Code → Deploy → Observe, Fix → Ship (workshop, with Sandhya Subramani). One agent, from a blank file to running in production before the session ended: build the loop, deploy serverless, wire up observability, break it on purpose, fix it with tracing, ship it. Session · Photos
- The Infinite Context Window Is a Myth: Context Engineering for AI Agents. Why stuffing everything into the window doesn’t scale, and what to do instead. Session
More photos from the week are here on LinkedIn.
I hope you enjoy this as much as I enjoy putting it together every week. It’s my little weekly de-stress moment. Let’s stay informed together and learn something new along the way.
See you next Friday,
Eli
Dev.to · LinkedIn · GitHub · Twitter/X · Instagram · YouTube