Stripped of buzzwords and jargon, an AI agent is a simple concept. It is a program that has 'knowledge' of multiple APIs connected to an LLM. The LLM makes decisions on how to call the APIs and how to interpret the results, usually in response to user input.

An obvious challenge follows from this: For an agentic future to exist, APIs must be simple to integrate. API discovery and documentation must thus be up to a standard where APIs can be trivially understood and executed by LLMs.

This is not currently the case with most API knowledge. The knowledge exists, but is piecemeal and often in disconnected silos or hidden by gatekeepers trying to extract a toll on other people's services. Standards exist but they are not always followed and are often ignored by vendors seeking to 'own' their own proprietary formats and increase lock-in.

I don't believe the solution to this problem lies in creating more proprietary formats or yet another paywalled silo. Public API knowledge needs to be open, free, community-driven, and accessible to all.

Today we launch the Open Agentic Knowledge repository (OAK), which I believe presents the 'real' solution for public API discovery and integration. It represents an effort to collate all of the world's public API knowledge into a single accessible source, and is a small-but-crucial first step toward a useful Agentic world.

Over the next few weeks, we'll be following up this initial launch with a few key tools and components designed to make the OAK repository more useful:

• A workflow runner, which is designed to automatically execute Arazzo workflows or OpenAPI operations as defined by the documents provided by the OAK repository.
• A runtime SDK, which can generate LLM tool definitions from OAK workflows and operations, and control their execution via the workflow runner.
• An OAK Index API server, which will allow humans and agents to dynamically discover and execute the workflows stored within the OAK repository.
• An MCP server/plugin that combines all of the above into your agentic development environment, vastly simplifying the process of integrating APIs into your agent.

In addition, we'll be adding a web interface into the Index API server, and provide more 'traditional' ways of integrating OAK knowledge into your projects.

We use open standards throughout and release all knowledge and tools (with the exception of services we must run internally) under the most permissive license available (MIT). All of these tools aim to be completely agnostic of your preferred agent builders or LLMs, and most can be used independently of any services Jentic provides, allowing you to use the knowledge within the OAK repository in whatever way you prefer.

We also realize that we're very much in the 'Wild West' of agentic development. We're in a phase where developers are freely experimenting with all of the new tools available to us, and it's an exciting time to be a developer. The move towards wider adoption, however, requires us to better understand issues like security, deployment, privacy, and compliance as they apply to the agentic future.

We'll be taking a phased approach to this. The initial versions of our tools will attempt to simplify authentication and token management, but require some manual manipulation of the agent's runtime environment. Long term, we'll be working on more integrated management solutions with the goal of simplifying and demystifying agent deployment for real-world applications. As the industry develops, we'll be continually improving the overall safety of agentic deployment, moving the concept of agents from the realm of 'experiments' into the realm of 'reliable production tools'.

We realize that the OAK repository is an ambitious project, and in its current state represents a long-term goal rather than a full-featured tool. We hope that you will join us in our journey, and that with every commit, every issue, and every pull request, we'll continue to build a more open and accessible agentic world.