If you’ve been following our series of blogs on AI in API management logs, you’ll have heard us talk about the Model Context Protocol (MCP) from Anthropic, and the Agent2Agent (A2A) protocol from Google with some excitement. Now it’s time to throw an important new standard into the mix: the Arazzo specification.
Standards are critical for successfully deriving value from the AI supply chain, and the fledgling Arazzo specification is no exception. Read on to discover all you need to know to start benefiting from it.
What is the Arazzo specification?
The Arazzo spec is a way to describe API workflows to complete complex tasks. Where an OpenAPI Specification describes the surface area of an API, the Arazzo spec defines how to complete specific workflows using APIs to achieve a business outcome. Think of OAS as a building block and Arazzo as a conveyor belt.
We’re proud to say that Tyk is partly responsible for helping with the delivery, launch, and advocacy of the Arazzo specification as part of our founding membership of the OpenAPI Initiative.
Why is the Arazzo specification important?
Interest in the capabilities and deployment of AI agents has exploded recently. This is why it’s so important to understand the value of the Arazzo spec. A quick recap for context:
- MCP enables AI clients (LLMs and AI agents) to easily discover and use tools to interact with their environments. It normalizes the interface between LLM client and tool with a meta-protocol that bypasses having to standardize LLM access APIs, instead normalizing tool and other context access.
- A2A enables AI agents to discover, interact, and collaborate with other AI agents; these agents might be using MCP to interact with the world.
Many modern enterprises will soon be running large sandbox environments of agents (some already are). These agents will work together through A2A, potentially building their own MCP tooling and deploying those to function-as-a-service (FaaS) environments to extend themselves for their task definitions.
Ultimately, this sandbox environment will be like a docker container – it will have carefully constructed access to the enterprise’s ecosystem of microservices, software-as-a-service (SaaS) applications, and integrations. This external access will most likely be encoded in MCP-provided tools and run through API gateways or sidecars for security and observability.
The benefits of Arazzo
What’s missing in the above scenario is a manual on how to complete complex tasks. This is where Arazzo comes in. Yes, you could use prompt engineering to tell your AI agents their workflows, but prompt engineering doesn’t deliver the benefits of standardization and automation that the Arazzo spec does. Arazzo provides a consistent and interoperable mechanism for AI models and agents to interact with APIs.
Agents don’t exist in isolation – they interact with tooling, which may or may not be built by AI-assisted engineers. That tooling will be running through a standard software development lifecycle (SDLC) that includes good quality assurance (QA) practices such as automated integration testing and continuous integration.
These QA pipelines are actually where most will generate their OpenAPI Specifications programmatically from their source code – these artefacts form the foundation of solid automation and API management practices. This is also where a diligent development team will generate Arazzo specifications that describe key workflows for their application, such as user enrollment, purchase lifecycles, and key business process automation. Why? Because the Arazzo spec is perfect for ensuring contract adherence in your API ecosystem for the wider business, before deploying anything to production.
If your SDLC is spitting out OAS and Arazzo documents, then these documents should be critical to ensuring AI agents behave themselves. Every time any critical APIs are updated, the OAS document will deliver those changes to the agent, and every time the workflow changes, the Arazzo spec should also be updated. Not through manual, flaky prompt engineering, but by easy-to-generate, automated and declarative documentation.
Embrace your AI-readiness
At Tyk we always saw the OAS specification as critical to API management success. It’s why we upgraded it from an export-artefact to a core management object of our internal API specification. It’s also why our AI Studio can import OpenAPI Specifications from files or directly from your Tyk Dashboard to generate tools for your LLMs to use. You can use these not only in our AI Studio managed chat rooms, but also in its built-in remote MCP server and local MCP-proxy server generator tooling.
We see the same future for Arazzo in the agentic landscape – declarative workflows, made available to agents to easily and accurately complete business processes in a precise and easy-to-test manner. So, don’t sleep on the spec. Consider how you can use Arazzo today to better define your business processes, and in turn get your organization AI-ready sooner.
Check out this handy article for further steps to AI-readiness.
