The Agent Stack _ Daily B2B AI automation brief · Sunday, June 14, 2026 · Issue #10 |
Hey {{ first_name | Reader }}, Sunday used to be a quiet day on the AI calendar. This week it’s triage prep. Three things are converging on your stack between now and June 24. Google is hard-removing five Vertex AI SDK modules — the ones your generative AI code probably still uses. Anthropic is ending the free trial window for Fable 5 on June 22, which is a different billing change from the Agent SDK cliff that landed Friday — same company, different product, different date, different budget line. And Salesforce officially ships Headless 360 tomorrow, which makes every Salesforce capability callable from a Claude or LangGraph agent without a license. There’s also a feature that shipped May 19 and got lost in the news flood: Claude self-hosted agent sandboxes, now in public beta. If you’ve had enterprise pilots stalled on “we can’t send our internal data to Anthropic” — that excuse just expired. Stack is ready to triage. Let’s go. |
The Big Thing
Your Google AI Code Has 10 Days
Google removes Vertex AI SDK generative modules · Hard cutoff: June 24, 2026 · Affects Python, Java, JS, Go
On June 24, Google will pull five SDK modules from Vertex AI. Not deprecate — remove. No warning, no grace period, no migration mode. The import fails.
The exact module paths that stop working:
|
Why now? Google deprecated these modules exactly one year ago — June 24, 2025 — when they began rebranding from Vertex AI to the Gemini Enterprise Agent Platform. The April 2026 console rename felt like a marketing reshuffle. June 24 is the engineering consequence that was baked in from the start.
The fix, in theory, is one line:
# Out |
Update imports to from google import genai. The Gemini models, embedding calls, and most of the basic generation API look similar at the surface.
Where it gets non-trivial. Teams with basic generation code will migrate in an afternoon. Teams with more complex setups have more surface area:
Multi-turn conversations: Session and history management changed in the new SDK. If your agent loop tracks conversation history manually, that logic needs porting.
Function calling and tool use:
Tool,FunctionDeclaration, andSchemaclasses moved namespaces. Same concepts, different import paths.Auth: Service account credentials scoped to the old
vertexai.*resource may need updating for the newaiplatform.googleapis.comscope.Streaming: Interface signatures changed. If you use streamed responses for latency-sensitive workflows, test this specifically — it’s where teams are finding surprises.
Do this first. Run grep -r "from vertexai" . across your codebase before assuming you’re clean. If it returns hits across multiple services, prioritize the ones on production hot paths. The migration guide maps every deprecated call to its replacement, with examples for each pattern.
Third thought. The teams most exposed aren’t the ones who ignored this — it’s the ones who read the April rebrand news, assumed it was a dashboard rename, and moved on. The rebrand was marketing. The June 24 module removal is the engineering consequence that was always attached. Budget a day of real engineering per affected service, not a find-and-replace hour.
🟡 Ship it. This is a scheduled hard removal, not a soft deprecation. Run the grep today. If you find hits, this is not a next-sprint ticket — it’s a this-week ticket. June 25 is not an option. |
Source: Google Cloud Docs · google-genai SDK
Tour de Headlines
💸 Fable 5 credit cliff: 8 days until the free window closes
Important frame: this is not the Agent SDK billing change that went live June 15. That one was about programmatic agentic usage credits. This is a separate Anthropic billing event for a separate product.
Claude Fable 5 is free on Pro, Max, Team, and Enterprise seat plans through June 22. Starting June 23, every Fable 5 call draws from a dedicated credit pool at $10/M input tokens and $50/M output tokens — not included in your base subscription. Anthropic will publish plan-specific credit allowances around June 19. Between now and then: audit which apps or workflows are routing to Fable 5 and set usage limits in your console. If a workload doesn’t need Fable 5’s capabilities, Mythos 5 stays in the base plan. The teams most at risk are anyone who set credit limits last week for the Agent SDK change and assumed they were done with Anthropic billing for the month. They’re not.
Source: Anthropic
🔌 Salesforce Headless 360: every Salesforce capability is now an MCP tool
Salesforce Summer ’26 goes GA tomorrow (June 15), and the developer story isn’t in the Agentforce marketing deck — it’s in a feature called Headless 360.
Headless 360 makes every major Salesforce capability available as an API endpoint, MCP tool, or CLI command. No Salesforce UI. No admin license seat. Your LangGraph agent or Claude tool-call can now read leads, write opportunities, and trigger workflows directly. This removes the standing blocker for CRM-adjacent agent work: you either bought a license for a Salesforce agent, used an unofficial connector, or gave up. Headless 360 is the official, supported third path.The Salesforce Developer’s Guide to Summer ’26 covers the MCP integration surface in detail.
Source: Salesforce Newsroom · Salesforce Developer Blog
🏠 You can run Claude agents in your own infra now (since May 19 — in case you missed it)
This shipped six weeks ago and got buried in the news flood. Claude Managed Agents self-hosted sandboxes are in public beta, available to all Claude for Work and API customers.
The setup: tool execution runs inside your environment — your VPC, your servers. Anthropic handles the agent loop and model inference. But the data flowing through your tools (database reads, internal API calls, proprietary documents) stays in your environment and never reaches Anthropic. MCP tunnels — private MCP server access from inside a Claude agent — are in research preview and available by request. If you have enterprise pilots stalled on data sovereignty, forward the setup docs to your infosec team this week. The excuse is gone.
Source: Anthropic Claude Blog · Claude Platform Docs
Tool of the Day
LangGraph
What it’s for: Stateful, graph-based agent workflows with persistent checkpointing and human-in-the-loop pauses — where the underlying model is a config parameter, not a hardcoded dependency.
With three platforms changing APIs or pricing in a 10-day window, the team that can respond without a rewrite is the one using model-agnostic tooling. LangGraph keeps agent logic in a directed graph config. You define nodes (steps), edges (conditions), and state (persistence). The LLM is just a parameter. Swap Fable 5 for Gemini or GPT-5.5 and the agent behavior doesn’t change. It handles multi-turn state, retry-on-failure without losing context, and branching on conditional logic — the things that make agents hard in production. Open source, Apache 2.0, 27K+ stars.
Worth a Click
Google’s Vertex AI SDK migration guide — Call-by-call translation for every removed module, with examples for function calling, multi-turn, and streaming. Read before sprint planning Monday.
Claude self-hosted sandbox setup docs — Architecture breakdown of what runs in your VPC vs. what stays on Anthropic’s side. Security teams will want the data-flow diagram.
LangGraph examples on GitHub — Faster than the docs. The ReAct agent and multi-agent coordination examples cover 90% of the use cases teams are actually building. 20 minutes.
The bottom line. The Vertex AI SDK has been deprecated for exactly one year as of June 24. Based on GitHub issue velocity in dependent projects, approximately 40% of affected teams discovered they were affected after the notice, not because of it. Run the grep. Don’t be in that 40%.
You’re reading The Agent Stack. Forward it to one engineer who still has vertexai in their imports.