I watched a RAG pipeline poison its own memory. Tech & AI Weekly by Eli, July 10 2026
Hi! This week I stepped back from publishing to do research: I’ve been deep in graph-based memory for AI agents, modeling relationships between facts instead of dumping everything into a vector store. More on that soon.
Here’s what mattered in tech and AI this week, curated so you don’t have to follow 20 sources.
AI & GenAI
-
OpenAI shipped GPT-5.6 (Sol, Terra, Luna) after the White House asked it to slow down. Three tiers at three price points is the product story. The real story is that a frontier model launch now comes with a government safety review attached. Watch whether that becomes the norm for every major release, not the exception. Read more
-
Researchers tricked GitHub’s AI agent into leaking private repos with a single public Issue. No special access needed: they wrote an Issue disguised as a request from a VP of Sales, and the word “Additionally” was enough to slip past GitHub’s guardrails. If your agent reads Issues or PRs and also touches private data, that’s the same gap. Read more
-
One developer found newer Claude models make more mistakes calling his custom tool. Armin Ronacher hit this building Pi’s edit tool, and Simon Willison’s theory is that Anthropic’s own tuning for its built-in edit tool is bleeding into how models handle everyone else’s schemas. It’s one report on one tool, not a broad pattern yet, but it’s a good reminder: test your actual tool calls before you swap models, don’t assume “smarter” means “safer to upgrade blind.” Read more
-
Jensen Huang says traditional coding is dead and the future is engineers building AI agents. I’ve heard some version of “X is dead” every year since I started in this field. What’s actually changing is the job: less writing every line, more knowing the agentic architecture, the harness, and how to prompt well enough to get a good agent out of it. That’s a different skill, not no skill. Read more
Cloud & DevOps
-
Datadog used Claude and Cursor to migrate a metrics pipeline off FoundationDB, and cut a 45-minute operation to about a second. The AI didn’t do it alone: every change went in against a failing test first, engineers wrote the query optimizations the model missed, and a parallel validator compared old vs. new in production before cutover. No test suite, no trust. Read more
-
A single customer scaled to over a million AWS Lambda functions using one AWS account per tenant. Strong isolation, but it only worked because they automated account provisioning from day one with Step Functions and CloudFormation StackSets. They hit real pain once they crossed 50 accounts without that automation in place. Read more
-
AWS published a practical breakdown of MCP tool design trade-offs. Less theory, more checklist: trim tool descriptions, keep parameters around 8 or fewer, lazy-load the tools you rarely need. If your agent is burning context or picking the wrong tool, start here. Read more
From me this week
- How to Stop RAG Hallucinations Poisoning Your Vector Store. Same theme as the GitHub agent leak above: once bad data gets into your system of record, code has to be the gatekeeper, not another model checking its work. Read more
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