Agents grow up: Google brings ADK to Go, while Claude spreads into Slack, Chrome, and your toolchain

If you've been treating "agents" as a fun prototype category-cool for a hackathon, risky for production-this week didn't let you keep that comfort. What caught my attention is how aggressively the industry is dragging agents out of the prompt playground and into real software surfaces: typed backend languages, team chat, the browser, and multimodal pipelines that look suspiciously like actual workflows.

Google's Agent Development Kit (ADK) showing up in Go is the clearest signal. Anthropic wiring Claude into Slack and expanding "Integrations" via MCP makes the distribution play obvious. And Microsoft Research pushing a Planner-Critic agent for long video and giant image sets says the next wave of agents won't just read text-they'll review footage, triage visual evidence, and still have to be correct.

Let's talk about what's really going on.

---

Main stories

Google ADK coming to Go is a bigger deal than it sounds

Go isn't the language you pick because you love fancy type systems or elegant metaprogramming. You pick it because you want services that don't fall over, concurrency that's straightforward, and teams that can ship. So when Google extends its open-source ADK to Go, I read it as: "We want agents to be built by backend teams, not just ML teams."

Here's what I noticed: a Go-native agent toolkit quietly reframes what "agent development" is supposed to look like. It's less "prompt + vibes" and more "workflows + evaluation + deployment." Strong typing matters here because agents aren't just generating text-they're choosing tools, calling APIs, passing structured arguments, and coordinating steps. Every one of those edges is where production systems break.

For developers, the "so what" is pretty immediate. If your infra is already Go-heavy (Kubernetes operators, internal services, data pipelines), you can now build agentic systems without forcing your team to hop into a Python-first ecosystem or stitch together five libraries that all disagree on abstractions. Concurrency is the other big point. A lot of agent patterns want parallel tool calls, speculative execution, and timeouts that don't turn into spaghetti. Go is built for that.

The competitive angle is fun too. Google isn't just shipping models; it's trying to own the way agents are built. Toolkits are distribution. The winner isn't necessarily the best model-it's whoever becomes the default wiring harness for "LLM + tools + evaluation + deployment."

Microsoft's MMCTAgent shows where multimodal agents are actually heading

Microsoft introduced a Planner-Critic framework (MMCTAgent) for reasoning across long videos and large image collections. That sounds academic until you realize how many businesses are sitting on mountains of visual data they can't query: security footage, retail cameras, industrial inspection images, medical imaging archives, sports clips, training videos, you name it.

The key idea-planner generates a structured approach, critic checks and refines-is basically an admission that raw "ask a model about a 2-hour video" isn't good enough. You need a system that can decompose the task, retrieve the right moments, call tools, verify intermediate outputs, and keep going without losing the plot.

This is interesting because it aligns with a pattern I'm seeing everywhere: agent architectures are turning into small organizations. One component proposes. Another component audits. Tools do the grunt work. Memory/retrieval keeps the state. It's less "one big brain" and more "workflow with guardrails."

If you're building products, the opportunity is obvious: multimodal search and investigation. But the catch is also obvious: cost and latency. Long-video reasoning is expensive, and teams will need strategies like sampling, scene detection, embeddings, caching, and "good enough" heuristics. The companies that win won't be the ones who can describe a fancy agent. They'll be the ones who can run it cheaply, quickly, and reliably-then wrap it in a UI that normal people can use.

Anthropic's Claude is playing the distribution game hard: Slack, Chrome, Integrations, and "Research"

Anthropic shipped a cluster of updates that all point in the same direction: Claude isn't just a chatbot tab anymore. It's trying to become a layer that sits inside the places work already happens.

Claude in Slack is the most straightforward move. Teams live in channels and threads. That's where decisions get made, context accumulates, and messy requirements get negotiated in real time. A model that can participate there-while respecting enterprise permissions-is immediately more useful than a separate app nobody checks.

I'm opinionated about this: Slack is one of the highest-leverage "agent surfaces" available. Not because Slack is magical, but because it's already the bus where status updates, links, screenshots, and small decisions flow. Drop an assistant into that bus and suddenly the assistant has ambient context. That's powerful. It's also dangerous if permissioning and leakage aren't airtight, so Anthropic emphasizing access controls is the right marketing line and, hopefully, real engineering.

Then there's "Claude Integrations" via MCP (Model Context Protocol). This is the more strategic piece. Integrations are how you turn a general model into a useful coworker: connect docs, tickets, repos, databases, dashboards. The moment you do that, the assistant stops guessing and starts citing. And Anthropic is pairing this with expanded "Research" features that pull from multiple sources and produce citations.

Here's the real significance: the industry is converging on a standard shape for enterprise AI. It's not "a model." It's an orchestrator with connectors, retrieval, tool calling, and some notion of traceability (citations, logs, evaluations). MCP is Anthropic's bet that interoperability wins. If MCP becomes the lingua franca, it reduces integration friction and makes Claude easier to adopt without rewriting your internal tool ecosystem.

Claude for Chrome (pilot) is about the nastiest problem: prompt injection in the wild

A browser assistant sounds like a simple UI change. It's not. The browser is where untrusted text lives. Every webpage becomes a potential attacker. So Anthropic piloting Claude in Chrome while explicitly focusing on prompt injection defenses is one of those updates that's easy to gloss over, but it's actually the hard part of "agents everywhere."

Prompt injection isn't theoretical. If your assistant can read a page and take actions, you've built a system that can be socially engineered at machine speed. The browser is the frontline.

What I want to see next (and what I don't know yet from this pilot) is how transparent the defenses are to developers and users. Do we get clear UI boundaries between "web content" and "trusted instructions"? Do we get audit logs? Do we get policy controls for actions? The teams that get this right will own a category. The teams that get it wrong will create the next enterprise security horror story.

Claude "Skills," prompt best practices, and Claude Code usage: the product is the workflow, not the model

Anthropic also published guidance on "Skills" for improving frontend design output (typography, themes, motion, backgrounds) and a set of prompt engineering best practices. Plus, they shared how their internal teams use Claude Code for navigation, testing, debugging, prototyping, docs, and automation.

On the surface, these are educational posts. Underneath, they're product strategy: standardize how people get good results. Skills are essentially packaged prompting + constraints + taste. And yes, that sounds soft, but it's exactly what teams need. The average dev doesn't want to become a prompting monk. They want a repeatable move that works.

The internal Claude Code write-up matters because it normalizes "AI as a daily tool," not an occasional assistant. If your org is still arguing about whether copilots help, Anthropic is basically saying: we're already using this across departments, and here's how. That's a cultural signal as much as a technical one.

And the new Max plan? It's the monetization tailwind behind all this. Higher limits and priority access aren't sexy, but they're how power users and teams stop bouncing off rate limits mid-project. Pricing is product. If the best workflow is "stay in the session, keep context, iterate," then the plan that enables longer sessions is a feature, not just a billing tier.

---

Quick hits

Microsoft Research pages briefly returning "high demand" placeholders is small but telling. Research blogs are now part of the hype and execution pipeline, and even the content layer is feeling load. It also made me wonder how many people are tracking "agentic security" topics like automated red-teaming for codegen-enough to spike demand, apparently.

---

Closing thought

Across Google, Microsoft, and Anthropic, I see the same pattern: the AI race is shifting from "who has the best model" to "who has the most deployable system." That means typed SDKs, evaluation loops, connectors, permissioning, and defenses against adversarial inputs-especially in places like Slack and the browser where real work (and real risk) lives.

If you're a developer or founder, my takeaway is simple: stop thinking of agents as a feature. Start treating them like a new application runtime. The winners will be the teams who can make that runtime boring-observable, secure, cost-controlled, and easy to integrate-while everyone else is still arguing about prompt templates.