Our yearly Summit was a great experience! The theme was evolving APIs for the agentic era, and we covered it from countless angles, on and off stage.

Nordic APIs Summit 2026 is a wrap! We experienced two full days (three if you count the UnConference) of deep API insights and themes around governing agents. If you attended, spoke, volunteered, sponsored, or participated in some other way, thank you so much for being involved in the event! You are what makes the Nordic APIs community so special.

In my opening keynote, I looked at how agentic AI is forcing a major shift in API interaction. Agents discover tools automatically, guess URL structures based on their training data, and retry in bursts. They’re nondeterministic and won’t always do what you tell them to or what you intended. I think this is a major shift for our industry, and one where the best practices are just now being established.

I ended my talk by posing an open-ended question: With all these shifts in mind, how do we prepare APIs for agentic consumption? Below, we’ll look at some of the solutions that our sessions and panel discussions covered. I’ll round up just some of the sessions I was able to catch — it’s by no means an exhaustive list.

WunderGraph and MaintainX keynote at Nordic APIs Summit.

Don’t Reinvent the API, Reinvent the Consumption Layer

No one recommended redesigning or reverse-engineering legacy APIs to shift into the AI agent era. Instead, the focus was on preparing how we expose what we already have to agents.

Lorna Mitchell, API Architect, TM Forum, recommends using established standards for this purpose. She suggests making your OpenAPI definition discoverable for agents, adding more field descriptions to improve agent usability, and using OpenAPI add-ons such as the Overlay Specification and the Arazzo Specification. “The standards are there, they are prior art, they are well known, and they will ground your agents,” she says.

For Kin Lane, maker of the infamous Kin Score on APIs.io, creating Model Context Protocol (MCP) servers and agent skills is a part of that, but they require additional best practices that go beyond pointing to an API and clicking “generate.” He made it clear that agents are already finding and integrating APIs, whether we like it or not. So, the response is important.

Our panel discussion on designing for the agentic experience highlighted other areas. Moderated by Gertrude Chilufya of Reframe Tech, panelists Erik Wilde of the OpenAPI Initiative, Kin Lane, and Daniel Kocot of adorsys recommended:

• Understanding the agent readiness of your API
• Improving descriptions, such as adding more error messages
• Using AI to probe around and find cracks in APIs
• Creating deterministic workflows for agents to consume

Our panel discussion on designing for the agentic experience.

Prepare Your Foundations Before Adding AI

Many sessions reiterated that preparing your foundations is much more important than jumping into wrapping agentic AI access around everything. So much boils down to your underlying data foundations. As Erik Wilde mentions in his presentation, it’s the core foundation — the APIs and the data strategy they touch — that matters most. If this isn’t structured well, what’s the point of APIs, anyway?

Furthermore, agentic AI exacerbates existing API security gaps, such as broken object-level authorization (BOLA), meaning the foundational API layer needs pruning to avoid pre-existing vulnerabilities and risks. The OWASP API Security Top 10, most recently updated in 2023, is still a widely cited resource for preparing this foundation.

Some speakers suggested dogfooding AI agents for this purpose — agents can aid not only functional testing but also penetration and chaos testing. This could expose gaps around unexpected inputs, as well as odd behavior that might break or exploit business logic.

Rob Allen goes through OWASP’s top ten list for APIs.

Put AI Governance at the API Layer

You can’t just tell AI “don’t touch production” in the prompt. At the Summit, we shared plenty of cases where agents can go haywire and disregard user instructions. As a result, agents need harder guardrails to limit destructive actions.

Author Jürgen Schulze painted a picture of this in his keynote by comparing agent-to-agent communication to a game of telephone, with messages losing fidelity from one share to the next. Jacob Ideskog of Curity also exemplified the new burdens in his opening Day 1 keynote, where he compared the shifting landscape to the changes his grandparent’s convenience storefront went through during the twentieth century.

Curity’s CTO Jacob Ideskog describes how the storefront has changed.

What solution do folks propose? It’s more than just cutting off access to important endpoints. For agents to be useful, they must be actionable. But we can’t treat agent identities and permissions in the same manner we dole out permissions to users.

In his closing keynote, Gravitee’s Sam Prodger put it well — you wouldn’t hire an employee without onboarding them properly. In reality, agents receive the same (if not more) access, often without much forethought.

For agents to be actionable (and safe), they require practices like just-in-time access and removing standing privileges. Curity’s Daniel Lindau showcased how permission escalation is quite real in agentic systems and why we need zero-trust checks at the API layer. Meanwhile, Curity’s Judith Kahrer explored specific standards, like OAuth and the MCP Authorization Specification, and how they interact with an identity server.

Optimize the Interaction Layer

One recurring theme was to keep things plain and simple. Your agent expects APIs to look and behave a certain way. If you give it humor, odd endpoint naming, or things left open to interpretation, or make off-the-cuff human-style remarks in API documentation or developer portals, this could lead to confusion.

There are other ways to optimize how agents interact with APIs and tools, too. One is to optimize the agent harness’s interaction with MCP. Hugo Guerrero from Solo.io demonstrated that simply removing unnecessary tool descriptions can cut token consumption tenfold or more. Other optimizations, like tool filtering, progressive disclosure, code mode, compression, and caching, can also reduce unnecessary token use.

In his closing keynote, Kin Lane described how his Kin Score evaluates many API interaction points to determine agentic readiness, including descriptions, idempotency, error semantics, rate limiting, MCP server quality, agent skills, and more.

For Shalev Shalit of Willow, it’s also a matter of determining the right interaction method (CLI, MCP, API, or browser) for the situation at hand. If it needs to work at scale for enterprises, he recommends MCP for access and auditing purposes. That said, ensure agents can access all types, he adds.

Solo.io’s Hugo Guerrero presents at Nordic APIs Summit.

New Infrastructure (May Be) Required

Our speakers also outlined areas where additional tooling may be required to evolve for the agentic era. For instance, Natalia Venditto from Adobe described how the AI harness — the infrastructure sitting around the large language model itself — is evolving. Others have recently made the case for a cloud-native agent harness.

Another way to reduce agent confusion is to segment which parts of a schema agents can see, since parsing a 200,000-token specification can be challenging. MaintainX took the stage to describe how it uses a router from WunderGraph to handle agent authentication and communication with subgraphs that clearly describe tools.

If you’re considering pricing your MCP servers, Josh Twist from Zuplo recommends metering and pricing per outcome, as opposed to metering per call or monetizing MCP distribution as part of an upsell. This may require custom logic within API or MCP gateways to track usage.

Natalia Venditto shares how the AI harness is shifting.

That’s a Wrap for Nordic APIs Summit 2026!

I was honored to moderate another Nordic APIs Summit. When I asked the crowd who was a repeat attendee, 30% or more must have raised their hands! I’m very impressed by the commitment to the event, and it’s a telling sign of what we’ve been able to build over the last ten years.

We had over 60 speakers across three stages and more than 300 total attendees at this Summit. I know many of you traveled from far away to be here, and we truly thank you for experiencing the event with us.