The Agent Stack, With Analysis
For the people building agentic systems — a daily read on what actually shipped across the agent stack: framework changelogs, cloud-platform announcements, and the press releases from the companies racing to own agents. Free. 5 days a week.
Sign up
Create your free account. We'll email you a link to verify — then you can comment on any analysis or prediction.
Recent issues
-
Industry story
AWS P6-B300 GPU instances expand to Seoul Region
Seoul getting AWS P6-B300 instances matters more than a routine regional rollout. These are the serious boxes: 8 NVIDIA Blackwell Ultra GPUs per instance, 2.1 TB of GPU memory, and 6.4 Tbps of interconnect, which is 2x the bandwidth of the previous generation. For teams training or serving trillion-parameter models in Asia Pacific, that memory ceiling and networking headroom are the constraints that actually bind.
Aug 24, 2026
-
Industry story
AWS Partner Central MCP Server adds OAuth via AWS Sign-In
AWS quietly closed a real friction point for partners building AI agents on top of Partner Central. Before this update, connecting tools like Amazon Q or Kiro to co-sell workflows, funding applications, and Marketplace seller setup required standing up a separate MCP proxy with SigV4 credentials. OAuth via AWS Sign-In collapses that into the IAM policies and CloudTrail audit trails partners already run.
Aug 24, 2026
-
Industry story
Google Antigravity Agentic Coding Tool Expands to Enterprise Gemini Subscriptions
Google bundling Antigravity into existing Gemini Enterprise subscriptions is the right move, and the governance layer is what makes it real. Spend thresholds, pooled token quotas, audit logging, MCP access policies from a central admin console: that's the checklist a CIO actually needs before saying yes to autonomous code execution at scale. The IDE coverage (VS Code, Visual Studio, JetBrains, Zed) removes the last friction point for developers.
Aug 24, 2026
-
Industry story
Pydantic AI v2.33.0 fixes Anthropic 1.0.0 compatibility breakage
Anthropic's SDK 1.0.0 dropped August 20 with a clean break: rebuilt on httpx2, legacy httpx support gone, and any Pydantic AI install that auto-resolved the new version started failing at runtime. Pydantic AI v2.33.0 patches it by pinning to anthropic>=1.0.0, but there's a catch for anyone passing a custom HTTP client to AnthropicProvider: you now need an httpx2.AsyncClient. The new SDK rejects legacy clients at construction time, so this isn't a silent degradation.
Aug 24, 2026
-
Industry story
Google Cloud Shares Agentic Vulnerability Discovery Harness Architecture
Google Cloud's release of the Agentic Vulnerability Discovery Harness architecture is the right move, and the timing matters. Offense has been using AI to find exploits faster than defenders could manually track them; AVDH is Google's attempt to close that gap on the defensive side, pairing autonomous multi-step code review with continuous scanning tools like CodeMender. Wiz Research is already showing what this looks like in practice, using multi-agent pipelines to reverse-engineer actively exploited CVEs, including Metabase CVE-2026-72898.
Aug 24, 2026
-
Industry story
Google DeepMind publishes "Intelligent AI Delegation" framework for multi-agent systems
Multi-agent AI systems have a problem nobody is talking about loudly enough: the longer the delegation chain, the more likely a subtly bad instruction gets accepted and passed down without question. Google DeepMind's "Intelligent AI Delegation" paper names that tendency the "zone of indifference" and builds four principles around it, covering how orchestrators decompose tasks, route to appropriately sized models, and limit data exposure via zero-knowledge proofs. The framework is framed for enterprise builders on Vertex AI, which means Google is also quietly making the case that its own stack is the right place to run agents designed to these standards.
Aug 24, 2026
-
Industry story
Google Cloud Deploys Agent-Based Security Review Pipeline for Product Launches
Google Cloud CISO Chris Betz just gave one of the clearest public looks at what agentic AI actually does inside a major enterprise: every product launch at Google Cloud now runs through an agent-based security pipeline, with high-risk flags routed automatically to human reviewers and threat models rebuilt as live dossiers instead of static documents. That last part matters most. When adversaries are generating malware mid-execution and running deepfake identity attacks, a threat model you wrote at launch is already stale before the product ships.
Aug 24, 2026
-
Industry story
Amazon Bedrock cuts GPT-5.6 Sol pricing 20–33% for agent workloads
Agentic model pricing is in freefall, and the operators who moved early on multi-step automation workflows are about to look very smart. OpenAI cut GPT-4.5 Sol API prices 20–33%, input tokens now at $4 per million and output at $20, and Amazon Bedrock is passing every dollar through to AWS customers at least through November 21. This follows earlier cuts to the Terra and Luna tiers, so the pattern is clear: the cost curve on autonomous coding agents and multi-step pipelines is dropping fast, and any budget case built on last quarter's numbers needs a refresh.
Aug 24, 2026
-
Industry story
Dragon Copilot Physician AI Agents Now GA on Microsoft Marketplace
Microsoft listing Dragon Copilot's physician AI agents on its own Marketplace is less a product launch than a distribution play. The capability existed; what changed is that U.S. health systems can now discover, evaluate, and purchase through a procurement channel their IT departments already use.
Aug 20, 2026
-
Industry story
Amazon Quick M365 Extensions GA: Agentic AI Inside Office Apps
Amazon Web Services has made its 'Amazon Quick' Microsoft 365 extensions generally available, bringing agentic AI (AI that can take multi-step actions autonomously) directly into Excel, PowerPoint, Word, and Outlook. The extensions pull context from users' own data and inboxes to perform complex tasks inside M365: redlining Word documents, building financial models in Excel, generating branded presentations in PowerPoint, and managing email and calendar workflows in Outlook. The launch positions Amazon Quick as a workplace agent operating inside existing productivity tools rather than as a standalone app, targeting finance, sales, marketing, legal, operations, and IT teams.
Aug 20, 2026