Agents Need Better Connectors, Not More Magic
2026-05-19 · 4 min read · Janaina Maia
The most important AI news is not always the loudest model launch. Sometimes it is infrastructure news that tells us where the product surface is moving next.
Anthropic acquired Stainless, a company that helps teams generate SDKs, command-line tools, and MCP servers. In plain English: Stainless helps software systems expose clean, reliable ways for developers and AI agents to use them. Anthropic’s own explanation was very clear: “agents are only as useful as what they can connect to.”
That sentence is worth sitting with.
The agent is not the product by itself.
For the last two years, a lot of AI product conversation has focused on the visible assistant: the chat window, the model, the prompt, the personality, the answer quality. That made sense when AI mostly responded to questions.
But agents are different. An agent is only useful if it can safely reach the right systems, understand the available actions, use the correct data, and leave enough evidence that a human can trust what happened. The intelligence in the model matters, but the connectors around it increasingly decide whether the experience is useful or reckless.
This is where I think many product teams are still under-designing. They are treating integrations as backend plumbing when they are actually part of the user experience.
Connectors are becoming trust surfaces.
An SDK, an API, or an MCP server sounds technical, but the design implication is very human: what can this agent see, what can it do, and how will the user know?
If an agent can connect to finance tools, customer records, internal documents, design systems, engineering workflows, or compliance evidence, the connection itself needs product design. Not just authentication. Not just a permission modal. A real work agreement.
- Which systems are in scope for this task?
- Which actions are read-only, and which can change something?
- What data did the agent use?
- What did it ignore?
- Where does the human approve, correct, or stop the work?
- What trail is left behind for review?
Those questions are not edge cases. They are the difference between an impressive demo and a product people can use inside serious work.
MCP is not just a developer protocol. It is an experience pattern.
MCP, or Model Context Protocol, is an open standard for connecting AI assistants to tools and data sources. The technical promise is interoperability: a more standard way for an assistant to access external systems without every integration being custom-built from scratch.
The product promise is bigger. If connectors become more standard, agents can move from isolated chat experiences into real workflows. But standard access also increases the need for standard expectations: visible permissions, understandable scopes, approval points, audit trails, recovery paths, and clear ownership.
In enterprise design, “the agent can do it” is never enough. The better question is: can the user understand and govern how the agent does it?
The design work moves closer to infrastructure.
This is uncomfortable for some design teams because it does not look like traditional UX work. It is not a new screen or a prettier chat interaction. It is product architecture, information boundaries, governance, workflow mapping, and operational trust.
But that is exactly why designers need to be involved. If connectors are designed only as technical capabilities, the human experience will become a black box. The user will see an output without understanding the route, the source, the authority, or the risk.
Good AI product design will make the agent’s reach legible. It will show where the model is acting from, which tools it touched, what assumptions it made, and where human judgement entered the loop.
My rule for agentic products.
Do not design the agent first and the connectors later. Design the operating boundary first.
Before asking how smart the agent should be, ask what it is allowed to access, what it is allowed to change, what evidence it must provide, and how a human can recover when something goes wrong.
Anthropic buying Stainless is not only a developer tooling story. It is a signal that the next phase of AI product design will be less about magic answers and more about trustworthy action.
The future of agents will not be won by the assistant that sounds the most confident. It will be won by the systems that connect carefully, act visibly, and make human accountability possible.