Hey there 👋 Two things happened that change where your agents live. Anthropic gave Claude a standing seat in your Slack — not a chat window, a coworker with a budget and an audit trail. And the Vertex AI SDK deadline we’ve flagged twice? It’s today. Your old Gemini imports are out of warranty as of this morning. The theme of the week holds: agents are autonomous actors that need a perimeter and a governor. This week they also got a desk. Five minutes — let’s go.
The Big ThingClaude Just Took a Seat in Your Slack — With Spend Caps and an Audit LogTL;DR: On June 23, Anthropic launched Claude Tag in beta — a persistent AI teammate that lives inside a Slack channel, summoned with @Claude. It replaces the old “Claude in Slack” app and is open to Enterprise and Team customers. The headline everyone wrote is “Claude is in Slack now.” The part that matters for builders is underneath: the identity, billing, and audit model got rebuilt for agents that act on a team’s behalf. What it actually does. Invite it into a channel and it keeps context over time, follows the conversation, breaks an assigned task into stages, and posts progress back into the thread where everyone can see it and pick up where someone left off. There’s an “ambient” mode that can proactively nudge people or flag follow-ups. It’s less a chatbot and more a standing member of the channel. Why this is the governance story, not the chatbot story. Usage moves off individual user accounts onto organizational billing. Each connected tool needs an admin-provisioned service account — not a person’s credentials. Admins set monthly token spend caps at both the org and the channel level, and every task Claude runs is written to a full audit log, including who asked for it. That is the exact perimeter our readers have been asking for all month, shipped as a product instead of a blog post about best practices. Third thought. The tell isn’t the feature list — it’s the timing. Anthropic says its own teams now route roughly 65% of code changes through an internal version of this. But Claude Tag launched the same day Claude had a major outage: 8,000-plus Downdetector reports, an elevated error rate across models, resolved around 12:44pm ET. So the honest read isn’t “agents are coworkers now.” It’s “agents are coworkers who call in sick — design for the day they’re out.” Ship it? — PILOT, don’t deploy wide. If you’re on Enterprise or Team, drop it into one channel with a channel-level token cap, provision a tightly scoped service account, and read the audit log for a week before you widen it. The legacy Slack app flips to Claude Tag on August 3 (admins opt in within 30 days), so you have runway. Just don’t put it on any path that has to keep running when Claude is down. Sources: VentureBeat · Search Engine Journal · TechRadar
Tour de Headlines⏰ Deadline day: your old Vertex Gemini imports die today. We called this at ten days out and again at migrate-now. Today is the actual removal date: Google deprecated the Vertex AI SDK generative modules exactly a year ago, and June 24, 2026 is when they go. Affected: vertexai.generative_models plus language_models, vision_models, tuning, and caching. Code that still imports them is expected to throw an ImportError. The 60-second check: grep -rn "vertexai.generative_models" across your repo. The fix is the google-genai package. Pinning an old version buys you days, not a fix — schedule the swap this week. Sources: Google Cloud migration guide · GCP Study Hub 💸 The agent budget is about to triple — and most teams still haven’t shipped one. Gartner pegs AI agent software spend at $206.5B in 2026, up from $86.4B in 2025, on the way to $376.3B in 2027 — the fastest-growing slice of enterprise software. Read the next line, though: only about 17% of organizations have actually deployed an agent, even as ~40% of enterprise apps are expected to embed task-specific agents by year-end. Gartner also warned in May that AI-driven layoffs free up budget but don’t, by themselves, deliver returns. The money is outrunning the execution — and that gap, between allocated and shipped, is where careful builders win the next twelve months. Sources: Gartner · Itential (Gartner Predicts 2026) 🪪 Give your agents identities, not secrets. The structural twin of Claude Tag’s service-account model: earlier this month (June 3) NetFoundry shipped zero-trust MCP and LLM gateways that give each agent a sovereign machine identity and broker its calls — so agents never hold API keys, service accounts, or shared secrets. It’s a vendor pitch, not gospel, but the pattern is the right one to evaluate after a month of credential-leak attacks: a leaked key only hurts if there’s a key to leak. If you’re wiring agents into real systems, the question for your next architecture review is “what identity does this agent have, and what brokers its access?” Source: NetFoundry (PR Newswire)
Sponsor Agents are getting service accounts and audit logs. Is your team getting better at the human part? RapportScore reads the human signals in every sales and customer call and scores how well your team actually connects — deterministic measurement, not vibes. Govern the agents; coach the humans. See where rapport breaks before the deal does. See your team’s score → |
Tool of the Day🧰 Google Gen AI SDK (google-genai)One client-based SDK for Gemini across Vertex AI and AI Studio — the thing you move to before today’s removal bites. Pairs with the deadline above. To be clear, this isn’t new — it’s GA and has been the recommended path for a while. Today just makes it non-optional. The shape is different from the old SDK: there’s no global vertexai.init(); you create a genai.Client (pass vertexai=True to point at Vertex), and model config moves to a per-request config object instead of constructor arguments. Response parsing for tool calls, safety ratings, and multi-part output shifts slightly, so budget an hour to update anything that reads those. One SDK now covers both Vertex and AI Studio — one code path instead of two. If you only do one thing today, do this. Read Google’s migration guide →
Worth a Click- Still vapor: GPT-5.6 and Gemini 3.5 Pro both miss the moment (Swisher Post · TestingCatalog)
As of today there’s still no official GPT-5.6 — GPT-5.5 remains current, a June 25 launch is rumored, and Polymarket’s window runs June 22–28. Gemini 3.5 Pro is still a limited Vertex-only preview (2M context and Deep Think confirmed; leaked pricing isn’t), with GA odds around 50–55% by June 30. The irony: Google removes a year-old SDK today while its flagship model still isn’t GA. Don’t roadmap on previews. - When the agents do the busywork, the office gets quiet (Fortune)
Anthropic’s own engineering lead for Claude Code and Cowork, Fiona Fung, told Lenny’s Podcast that leaning hard on agentic AI made the work “a lonely experience” — productivity up, human contact down — enough that the company started adding team-building to counter it. A real second-order effect worth a thought before you go all-in.
The agent got a desk this week. Google turned a year-old SDK into a pumpkin this morning, Anthropic gave Claude a standing seat in your Slack hours after it fell over for 8,000 people, and Gartner says the budget for all of it is about to triple while most teams still haven’t shipped one. The model and the SDK are the swappable parts. The spine you own — identity, spend caps, an audit log, a fallback, and the habit of migrating before the deadline instead of after — is what survives every model swap and every outage. A busy launch week isn’t a test of your FOMO. It’s a test of your governance. Stay sharp — The Agent Stack
Your daily 5-minute brief on AI agents, agentic workflows, and the automation tools B2B builders actually ship. Published weekday mornings by Pixiu Media Holdings LLC. | 
