AI Agents are a hot topic, with many parties giving different meaning to the definition. While I don’t think that there’s a right or a wrong definition, this is why we are quite excited about the Agent paradigm.
In computer science, a software agent is a computer program that acts for a user or another program in a relationship of agency.
Wikipedia
AI twitter on the other hand, generally uses the term AI agents for LLM powered programs that can autonomously accomplish complex goals.
Over the last 6 months, we have built AI agents for document summarisation, content generation, personal branding, code review and many other use cases. Our key realisation was that for complex use cases, a team of specialised agents is more reliable and useful than a single general purpose agent.
This is why, we started working on a simple architecture that defines a standard approach to an AI Agent API. Each agent is an autonomous application (in all effects a microservice) that can perform specific tasks with different levels of autonomy.
An agent can engage directly with a user through a UI, or can engage with other agents to help complete tasks. The level of autonomy can vary: it can be limited to using AI to complete a very specific task; it can be part of a pre-defined chain of actions that contribute to complete a job; or it can be emergent by an automated decision making process that allows agents to dynamically decide which route (and which other agents to engage) to reach a solution.
In all these cases users can simply delegate a task and go do something more useful (or enjoy a cup of tea) while agents are at work for them.
Our API protocol describe a discovery phase, where agents can introduce themselves, describe what tasks they are capable of and what kind of payload they can receive and return at the end of their job. The beauty of working with LLMs is that this exchange happens in plain English and it works pretty well both for humans and AI. Once humans or AI tools have discovered the capabilities of other agents, they can easily engage with them to deliver the services needed.
Having defined a first version of our A2A (agent to agent) api protocol, we are working on the construction of an initial set of agents with different skills: from extracting and analysing different types of data, to storing and manipulating information, to presenting results in the most readable and accessible format.
Our aim is to create a collection of agents, some with generic skills, others with very specialised ones, which can easily be dropped in our clients organisations to collaborate with their human colleagues with the goal to accelerate workflows and improve productivity.
We have been testing our architecture for several months and we have already deployed the first agents in a few organisations. Our plan is to soon publish the specifications of our API and release the first version of our platform to allow customers to engage with our agents in the first quarter of 2024.
Leave a Reply