🛠️ This Week's Deep Dive: Your AI Can Talk. Here's How to Give It Hands.
I've spent the last year watching the AI world argue over which model is smartest — GPT-5 versus Claude versus Gemini versus whatever dropped last Tuesday. The benchmarks swap leaders every month. The blog posts pile up.
Meanwhile, the actually-interesting thing has been happening quietly in the background: a whole new layer of the AI stack has shown up, and almost nobody is talking about it.
It's the layer that decides whether your AI can do anything — or whether it just talks to you about what it would do, if only it had hands.
This week I wired one of these platforms — Composio — into a real chatbot and watched a GPT-4o-mini agent create a contact in a live HubSpot account in about three seconds. Here's what I learned, and why I think the integration layer is now more important than the model you pick.
———————————————————————————
The Three Layers Nobody Draws on the Whiteboard
Every useful AI agent is actually three things stacked together:
1. Brains. The LLM. This is the part everyone talks about. ChatGPT, Claude, Gemini. The thing that turns your question into words.
2. Memory. A vector store or a retrieval system that lets the agent remember what it knows about your business, your docs, your past conversations. Pinecone, Chroma, your own Postgres.
3. Hands. The integration layer. The code that lets the agent do something in the real world — create a Slack message, book a calendar slot, write to your CRM, send an email, update a spreadsheet.
Most people — including most people building AI tools — spend 90% of their attention on the Brains layer. Which model is cheapest, fastest, most accurate?
But here's the thing: the brain of a very smart person who can't move their arms is still a person sitting in a chair.
An AI that can explain how to book a meeting is a chat toy. An AI that actually books the meeting is a product.
The gap between those two things is the Hands layer.
———————————————————————————
What Composio Is (And What It Isn't)
Composio is an integration platform built specifically for AI agents. Think of it as "Plaid for AI" — it handles the 1,000 annoying little things between your agent and the 250+ SaaS apps it might want to touch.
Concretely, it does three jobs:
OAuth handling. You don't write auth code for Gmail, HubSpot, Slack, Notion, Stripe, etc. Composio handles the OAuth dance once, stores the tokens, and your agent just calls tools.
Typed tool schemas. Your LLM gets a clean, machine-readable description of every tool — so when GPT-4o-mini wants to create a HubSpot contact, it knows exactly what fields are required.
Execution runtime. When your agent picks a tool and generates arguments, Composio actually runs the call, handles retries, and returns a structured result.
It's not a competitor to LangChain or CrewAI — it's the tool layer beneath those frameworks. You bring your own agent (LangChain, raw OpenAI SDK, whatever). Composio brings the hands.
Pricing starts free and scales usage-based from $29/month. If you ship agents for a living, it's the fastest time-to-first-integration in the category right now.
———————————————————————————
The Live Build: HubSpot Write-Through in One Afternoon
I wanted to see if this was real, so I took the chatbot I'd already built for a client demo and tried to add a CRM integration.
Baseline: a GPT-4o-mini chatbot that qualifies leads and stores them in an in-memory dashboard. Nice, but when the demo ends, the leads live in a CSV.
Upgrade goal: every captured lead flows directly into a real HubSpot CRM in real time, with name, email, phone, project type, and a custom "lead source" field.
Here's the entire integration, trimmed to the interesting lines:
from composio import Composio
from composio_openai import OpenAIProvider
from openai import OpenAI
composio = Composio(provider=OpenAIProvider())
tools = composio.tools.get(
user_id="[email protected]",
toolkits=["HUBSPOT"],
)
response = openai.chat.completions.create(
model="gpt-4o-mini",
messages=history,
tools=tools, # <-- Composio's already formatted them
tool_choice="auto",
)
if response.choices[0].message.tool_calls:
composio.provider.handle_tool_calls(
user_id="[email protected]",
response=response,
)That's it. Eleven lines.
The agent now decides on its own when to call HUBSPOT_CREATE_CONTACT, maps the conversation fields into HubSpot's contact schema, and writes. I fed it a fake prospect — "Hi I'm Sarah Chen from Acme Supply, 602-555-0123…" — and watched a Sarah Chen contact appear in HubSpot's Contacts view about two seconds later.
A contact created by the bot, live in HubSpot.
It took longer to make coffee than to wire HubSpot in.
Full repo and the client-facing SKUs behind this build are on the CoffeyAI site.
———————————————————————————
The Honest Comparison
Composio isn't the only game in town. Here's how the main options stack up if you're building agents right now:
Approach | Strength | Weakness | Best for |
|---|---|---|---|
Raw API calls | Maximum control, zero lock-in | You maintain auth, retries, and schema forever | 1–2 stable APIs |
Zapier / Make AI | Non-dev friendly, huge app catalog | Slow, expensive per action, brittle in agent loops | Business users |
Composio | 250+ apps, managed OAuth, typed tools | Another vendor in the stack, usage pricing | Developers shipping agents |
MCP (direct) | Open protocol, no vendor lock-in | Fewer apps, you host the servers yourself | Long-term architecture |
A few honest trade-offs worth naming:
Composio adds a dependency. If they change pricing or get acquired, you'll feel it. The escape hatch is that tool calls are standardized enough to port.
Zapier AI is tempting if you don't code, but most agent loops involving multiple steps feel sluggish and expensive inside Zapier.
Raw API is great for one or two integrations. The moment you need three or more, the auth/retry/error-handling code becomes a second job.
MCP (Model Context Protocol) is the open standard everyone is converging on. Composio speaks MCP too. If you're making architectural decisions you'll live with for three years, lean toward MCP-native tools.
———————————————————————————
Who Should Adopt One This Month
Honestly?
Founders building AI-first products that need to touch customer systems. Composio or similar saves you three to five engineer-weeks on the first three integrations.
Freelancers and agencies shipping chatbots and internal tools for clients. The "your bot writes to your CRM" pitch moves projects from $500 one-off bots to $2,500+ Connected builds. The economics change completely.
Teams past the prototype stage whose agents are doing work worth tracking, with enough scale that the $29 starter tier is rounding error.
Who should wait: if you're still figuring out whether AI agents solve your problem at all, don't bolt on an integration layer. Use raw API calls for one clear use case, validate the workflow, then upgrade.
———————————————————————————
The bottom line: The Brains layer is commoditizing. The Memory layer is standardizing. The Hands layer is where the next 18 months of differentiation will live — and most people aren't watching it yet. That's the opportunity.
———————————————————————————
⚡ Quick Hits
1. Shopify's AI Toolkit is the "Hands" thesis in action — Shopify shipped a free, open-source plugin on April 9 that lets Claude Code, Cursor, Codex, Gemini CLI, and VS Code manage your store directly — products, inventory, apps, the whole operation — in plain English. Not through the dashboard. Not through a chatbot. Directly. This is what happens when a major platform decides the integration layer matters more than the UI.
2. OpenAI's GPT-5.4 can now run your computer — GPT-5.4 is the first general-purpose OpenAI model with native computer-use built in, not bolted on. Agents can now operate real applications across workflows. Combined with the tool-layer platforms discussed above, this is another signal that agent-first products — not chat-first products — will define the next 12 months.
3. Meta quietly abandoned open source — Muse Spark, Meta's first AI model under new Chief AI Officer Alexandr Wang, launched April 8 — and it's fully proprietary. No weights. No fine-tuning access. No community forks. After years of Llama being the open-source flag-bearer, the competitive pressure finally broke them. If open-source foundation models matter to your stack, the torch has passed to Mistral, Qwen, and Ai2.
4. The real cost of AI, in three numbers — Stanford's 2026 AI Index Report dropped April 13, and the environmental data is hard to ignore. Training Grok 4 alone emitted an estimated 72,816 tons of CO2. AI data centers now pull 29.6 GW combined — roughly what New York State uses at peak. And annual GPT-4o inference water use may exceed the drinking water needs of 12 million people. The frontier is real. So are the externalities.
———————————————————————————
💡 Prompt of the Week
The "Hands Inventory" Prompt
If you're planning to build — or hire someone to build — an AI agent for your business, run this prompt first to make sure the "hands" you'll need actually exist:
"I want to build an AI agent that handles [specific task] for my business. Walk me through every external system or account it would need to touch to do this job well — CRM, calendar, email, messaging, files, payments, anything. For each one, tell me (1) whether Composio, Zapier, or direct API is the best fit, and (2) any authentication or permission hurdles I should expect. Be concrete — name actual products."
This saves you from the most common agent project failure: building the Brains and then discovering there's no realistic path to the Hands.
———————————————————————————
🛒 Before You Go
If you want custom AI chatbots built for your business — the kind that actually write to your CRM, book your calendar, and ping your team on Slack — that's exactly what I build at CoffeyAI.
And if you're a DIY type, AI Prompt Power Pack has 65+ battle-tested prompts organized by use case:
———————————————————————————
One ask this week: What app would you most want your AI to connect to? Hit reply and tell me. I'm building the Issue #5 topic list from your answers.
———————————————————————————
Enjoying Wired on AI? This newsletter is built on beehiiv — the best platform for newsletters. If you're thinking about starting your own, it's where I'd start.
———————————————————————————
See you next week.
— Scott