Every agent decision needs a receipt. Tech & AI Weekly by Eli, July 17 2026
Hi! Friday again. I’m heads-down on two new series. One on agent memory done properly: JSON, vector, graph, and the hybrid graph plus vector setup I keep reaching for, with techniques to keep memory clean against hallucinations and prompt injection. The first memory post drops next week. And a separate one on agentic observability. Meanwhile, this week’s theme wrote itself: agents are getting deployed faster than anyone can tell what they’re actually doing, or spending. Here’s what I read so you don’t have to.
AI & GenAI
- Stripe published a benchmark of where AI agents fail. Finally, hard numbers on the failure modes we’ve been arguing about in slides. Read this before your next agent design review. Read more
- Agents are running up bills faster than humans can approve them. InfoQ walks through a case where an agent hit the cloud bill at superhuman speed and nobody had a guardrail. If your agent can spend money, this is your homework. Read more
- Every agent decision needs an “evidence packet”. The idea: don’t just log outputs, log the reasoning trace as an auditable artifact. This is where agent observability is going, and I think it’s the right call. Read more
- GPT-5.6 (Sol, Terra, Luna) is GA in Amazon Bedrock, and one team already migrated in prod: 2.2x faster, 27% cheaper. Both links belong together: the GA announcement and the real migration numbers. Do your own benchmark before you migrate.
Cloud & DevOps
- Google pitches “Agent Substrate” as the Kubernetes for AI agents. Kubernetes did it for containers, and now the fight is over who owns the runtime for the next decade. Direct pressure on where Bedrock AgentCore is going. Read more
- CNCF put out a report on cloud-native as the base layer for trustworthy agentic AI. Reads like the reference architecture I keep drawing on whiteboards, but with citations. Useful if you’re the one arguing for the boring platform work. Read more
- Azure quietly built an AI that decides when Azure itself is down. AI ops watching AI, and the design write-up is more interesting than the headline. Read more
From me this week
- How to Stop RAG Hallucinations Poisoning Your Vector Store. If your RAG pipeline is quietly writing bad data back to itself, this is the fix I wish I’d had earlier. Validate before you write. Read more
Where I’ll be next week
Flying to New York to meet my team in person. My team is scattered across the world: Mike in Australia, Dennis in Germany, the rest of us across the US. Getting everyone in the same room is rare, and it’s the best part of the job. New ideas always come out of those days, so expect some of them to land in future editions.
Where to find me
- Hack the Video Agent Context Graph. July 30, AWS Builder Loft San Francisco. A one-day hackathon I’m running: build a video agent that turns raw footage into a Neo4j graph it can reason over, using Strands, TwelveLabs, and OpenAI. Come say hi if you’re in SF. Details
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