Merlin Agent Builder
Building Agents on Tray's AI-ready iPaaS unlocks a whole new paradigm of automation.
We have accelerators that can help you get started in minutes or you can build your agent one tool at a time to meet your exact business requirements. With Tray the hardest parts of building agents (agent architecture, setting up function calling, defining tool schemas) are all simplified in our low-code interface. We make it easy to get started and leave to door open to endless customization.
This feature is currently being rolled out incrementally and may not yet be available to all users. Full availability is expected soon. If you're interested in using this feature, please reach out to your Customer Success Manager or Account Executive.
What is an AI Agent? Copy
An artificial intelligence (AI) agent is a software program that can interact with its environment, collect data, and use the data to perform self-determined tasks to meet predetermined goals.
If you are familiar with building on the Tray platform this is how these pieces break down:
Environment
Where is the agent being called from?
For chat interfaces, who is the person talking with the agent?
Collect Data
Real time business data
Knowledge resources specific to your business
Perform tasks
Run workflows that can reference other systems
Run workflows that take actions in other systems
Using an AI-ready iPaaS is much different from an agent that exists within a single platform because you can orchestrate tasks across tools that just isn't possible with an agent that is silo'd within a specific application.
Components of an Agent on TrayCopy
Merlin Agent BuilderCopy
This is the part of a project where you build and manage your agent. You can add or remove tools from the agent, rename your agent, and update agent tool descriptions.
Agent WorkflowCopy
Each agent is automatically provisioned with an agent workflow. Inside this workflow there are specific connectors (ai-agent-utils) that use AI to understand a request, reason about what should be done, determine what tools are available, choose the right tools, execute the tools, and respond with the outcome. The architecture of the default agent workflow is built to run without modifications (which includes session handling). We have default inputs for the agent, but you can customize these by updating the input schema of this workflow.
Agent System Prompt Copy
The "Run Agent" step inside the Agent workflow is where you provide the instructions for the agent at a general level, this typically includes Standard Operating procedures or routines that you have in mind that the agent should run through in natural language.
Agent ToolsCopy
Agents must have tools assigned to them to run successfully. Tools are Tray workflows that the agent can use to execute a request.
We like to think of tools in two ways:
Tools that provide more context to the agent
Tools that take actions in your systems
You can build tools that are used for Retrieval Augmented Generation (RAG) in order for the agent to have more knowledge and better answer requests as in the case of a Knowledge Agent accelerator. You can also build tools that take actions to help get work done like in the case of the IT Ticketing Agent accelerator.
Each tool is defined for the agent so it knows what the tool does and how to interact with it. You can update the tool definition by updating the inputs and outputs of the tools AND by adding a natural language description of the tool (this is handled at the level of the Tray workflow and can be modified from within the Merlin Agent Builder UI)
Invoking and Executing an AgentCopy
You can call an agent using the AI-Agent connector from any other workflow. We have guides and templates to setup an agent that can be deployed directly into your enterprise chat application (ie Slack). Other examples of invoking are when a new ticket is submitted to your support desk (ie Zendesk) or your IT service desk (ie Jira).
Agent LimitsCopy
Agents use Tray's AI infrastructure, you can read more about how Merlin Agent Builder uses your data here: LINK
Agents are limited to 50 concurrent runs within a single organization. If this limit is reached the agent steps within your Agent workflow will return a 429 temporarily.
You cannot move agents between projects or workspaces - we recommend creating your agent in a dedicated workspace so you can give other builders view access to the workspace so they can learn from what you build!