Last week, I ran a small experiment.
I wanted to see how well GPT-4 could interact with a real-world API without much hand-holding. Nothing complicated. No multi-agent workflows. No orchestration frameworks. Just a simple task that thousands of applications perform every day.
Send an email through SendGrid.
The goal was straightforward. Give the model the context it needed, let it generate the request, and see how far it could get.
What happened next surprised me.
GPT-4 generated a request containing several parameters that looked completely valid. The payload was structured correctly. The field names were descriptive. Everything looked professional.
The only problem was that those parameters did not exist in the SendGrid API.
The request failed immediately.
At first, I thought this was a model problem. After all, hallucinations are a well-known limitation of language models. **But the more I experimented with APIs, agents, and production workflows, the more I realized something deeper.
The problem is not that AI agents occasionally hallucinate APIs.**
The problem is that most APIs were never designed for AI agents in the first place.
For the last two decades, APIs have been designed around a very specific assumption.
The consumer is a developer.
That developer reads documentation, understands business context, interprets ambiguous descriptions, and fills in gaps when documentation is incomplete.
When an API specification says:
{
"status": 1
}
a developer can usually figure out what that means.
Eventually, they learn that:
1 = Pending
2 = Approved
3 = Rejected
and move on.
AI agents don't work that way.
They do not infer intent from tribal knowledge. They do not ask the developer sitting next to them for clarification. They only know what exists inside the contract they were given.
If the meaning is not explicit, the agent is left guessing.
And guessing is where things start to break.
What made the SendGrid experiment interesting wasn't that GPT generated an invalid request.
It was how convincing the invalid request looked.
The generated payload contained fields like:
{
"recipient_email": "john@example.com",
"email_subject": "Welcome",
"priority_level": "high"
}
None of those fields exist in SendGrid.
Yet if you've worked with enough APIs, they feel completely reasonable.
That's because GPT wasn't retrieving the schema.
It was predicting the schema.
Across millions of code examples, SDKs, documentation pages, and tutorials, fields like recipient_email and email_subject are statistically common. The model generated what seemed likely to exist.
The API, however, only cares about what actually exists.
This distinction is easy to overlook, but it sits at the center of many agent failures.
Language models operate on probability.
APIs operate on contracts.
Those are fundamentally different systems.
Historically, this wasn't a major issue.
A developer chooses an API once, integrates it into an application, and that integration remains relatively stable.
Agents change that model entirely.
Instead of discovering APIs during development, agents increasingly discover and use capabilities at runtime.
That sounds simple until you look at the scale of modern enterprises.
Large organizations often operate tens of thousands of APIs and hundreds of thousands of endpoints. Most engineering teams don't even have an accurate inventory of everything that exists.
For a developer, that complexity is hidden because someone already made the integration decision.
For an agent, the discovery process becomes part of the workflow itself.
The challenge is no longer "Can the API perform this action?"
The challenge becomes "Can the agent find the correct capability among thousands of possibilities and understand how to use it correctly?"
That's a very different problem.
One of the most interesting ideas I've come across recently is that enterprise APIs expose too much implementation detail and not enough intent.
Imagine a workflow that creates a new customer.
From a business perspective, that's a single action.
From an API perspective, it might require:
A developer can understand how those pieces fit together.
An agent sees five independent endpoints and must figure out how they relate to one another.
As API landscapes grow, this becomes increasingly difficult.
The problem isn't that agents lack intelligence.
The problem is that we're asking them to navigate systems that were optimized for flexibility rather than clarity.
The more I think about agent infrastructure, the more convinced I become that agents should interact with capabilities, not endpoint catalogs.
A business action like "Create Customer" should look like a business action.
Not a sequence of fifteen API calls hidden behind documentation.
Better API design will help.
Better specifications will help.
Better documentation will help.
But they don't solve the entire problem.
Even if an agent perfectly understands an API, production systems introduce an entirely different set of challenges.
None of these problems are reasoning problems. They're execution problems.
And execution is where many agent architectures still struggle today.
Most diagrams describing AI agents look something like this:
LLM → API
In practice, production systems need something in the middle.
An execution layer.
A layer responsible for authentication, validation, retries, observability, and policy enforcement.
The model decides what it wants to do.
The execution layer determines whether that action can be performed safely and reliably.
Without that layer, every API call becomes a potential point of failure.
The model is forced to handle responsibilities it was never designed for.
And reliability quickly becomes difficult to achieve.
While building agent workflows, we kept running into the same pattern. The model wasn't struggling to decide what action to take. It was struggling with everything that happened after the decision was made.
The more integrations we connected, the more obvious it became that agents needed infrastructure around API execution, not just better prompts.
That realization eventually became one of the motivations behind Swytchcode.
Instead of treating APIs as raw endpoints that agents need to figure out at runtime, we started treating them as structured capabilities with managed execution underneath. The goal wasn't to make the model smarter. It was to make execution more reliable.
The phrase "execution layer" can sound abstract, so let's make it concrete.
Imagine an agent wants to create a customer in HubSpot, send a welcome email through SendGrid, and post a notification to Slack.
From the model's perspective, those are simple actions.
But behind the scenes, each integration comes with its own set of requirements.
In many agent architectures today, the model is expected to handle all of that complexity directly.
That's where things start to break.
What we've found is that agents work much more reliably when API execution is treated as infrastructure rather than prompt engineering.
That's one of the ideas behind Swytchcode.
Instead of exposing raw APIs to agents, Swytchcode provides a managed execution layer that sits between the agent and external services.
That layer handles things like:
As a result, the agent can focus on intent:
Create a customer.
Send an email.
Update a CRM record.
The goal isn't to replace the model.
The goal is to provide the infrastructure that allows the model to operate reliably in production.