Ran Microsoft's Cosmos DB Agent Kit (100+ best-practice rules) against my own sample repo. Caught a real bug: mock embeddings never normalized to unit length, silently breaking cosine-similarity scores.
sjwiggers.com/2026/09/29/cos…#agenticai
Apple Reference Image signs pixel data inside the iPhone 18 Pro sensor, then develops it in Private Cloud Compute under Apple's signature.
Trust moves from C2PA's editing chain to Apple. Verification stays on Apple platforms.
infoq.com/news/2026/09/apple…
A "10 per second" rate limit let 19 through in 200 ms. Add a second replica and every in-process limiter doubled. I load-tested six rate limiting algorithms on Azure, so you don't have to guess.
sjwiggers.com/2026/09/24/rat…#ratelimiting#azure
1,727 price meters for Foundry Models in one Azure region. None carries a model id. So I built a ledger that joins the ARM catalog, Retail Prices and your own deployments: cost per 1M tokens, retirement dates, what breaks.
sjwiggers.com/2026/09/22/fou…#microsoftfoundry#azure
AWS says it cannot restore data held exclusively in a damaged UAE availability zone, or exclusively in its Bahrain region, after strikes during the conflict with Iran.
Multi-AZ and S3 durability are regional properties.
infoq.com/news/2026/09/aws-m…
Same user, session revoked. CAE-capable client lost access in 4 seconds. The other kept going for 65 minutes.
Six deployable samples measuring what the Entra ID architecture diagram compresses:
sjwiggers.com/2026/09/20/mic…#entraid#conditionalaccess
Real production case: an account at 100% RU utilization, throttling under load. Cause wasn't throughput — one partition absorbing 80% of traffic. Fixing the model dropped utilization to 20-35%.
sjwiggers.com/2026/09/17/cos…
GPT-6 Astra is the first model OpenAI classifies as Critical for cybersecurity under its Preparedness Framework.
In expert-led testing it found unknown vulnerabilities in a browser and an OS kernel and built working exploits.
infoq.com/news/2026/09/gpt-6…
Foundry Agent Service BYO thread storage: Cosmos DB needs 3,000 RU/s minimum or you get CapabilityHostProvisioningFailed. Classic and New runtimes also write to different containers entirely.
sjwiggers.com/2026/09/16/cos…#agenticai
Azure Virtual Desktop Hybrid is GA: session hosts on your hardware via Arc, brokering in Azure.
Windows multi-session isn't supported, so the density model AVD is sold on doesn't apply. The Hybrid service license is still unpriced.
infoq.com/news/2026/09/avd-h…
Built all 5 RAG architectures (Hybrid, GraphRAG, Agentic, Corrective, Multimodal) on the same Azure infra, same 15 questions. Weirdest finding: Corrective RAG scored better on questions it wasn't built for than the one it was. sjwiggers.com/2026/09/13/fiv…#rag#azure
MCP vs API is the wrong question. I put both in one Azure Functions app: same backend, a REST door for developers, an MCP door for agents. The agent chained three tools nobody orchestrated. Full sample, Bicep, and deploy lessons:
dev.to/steefjan_wiggers_34a4…#azurefunctions#mcp
47,004 input tokens for a 280-token answer. My scam-checking agent on Azure Functions hosted skills taught me when NOT to let the model orchestrate: gather evidence in code, keep judgment in the model. Full story + repo:
dev.to/steefjan_wiggers_34a4…#azurefunctions#aiagents
Multi-agent systems need two memory scopes: shared state for coordination and private state to keep each agent's persona separate. Handoff mechanism: change feed. One agent's write is the next agent's signal; no polling. @AzureCosmosDBsjwiggers.com/2026/09/10/cos…
Ran the same question 4 ways against one Cosmos DB container: recency, semantic, hybrid, keyword. Four different answers — "find the right memory" isn't one query pattern.
sjwiggers.com/2026/09/03/cos…#agenticai
What happens when you actually build nine agentic AI design patterns on Azure instead of just diagramming them?
Steef-Jan Wiggers shares what worked, what broke, and lessons around Azure Cosmos DB RBAC, Foundry, Event Grid, Bicep, and more.
msft.it/6011aTZ8W#AzureCosmosDB#AgenticAI
deselect at consent. MCP servers are the motivating case.
Partial consent isn't new (GitHub, Google, Entra all have it). Developer control over what's droppable is.
infoq.com/news/2026/09/cloud…
Partition key and item shape decide whether a Cosmos DB agent memory schema holds up.
TTL trick: -1 on the container, override per item. Short-term turns expire; long-term memories don't. Same container.
sjwiggers.com/2026/09/01/cos…#agenticai
Agent ≠ Model. Foundry's model router now spans 28 regions and refreshes its pool under a stable endpoint. Models are becoming swappable parts; the durable investment is the control plane. Where routing is the wrong answer: sjwiggers.com/2026/08/31/bui…#microsoftfoundry