Building Your First Agent
Agents are the chat interface that can tie all your complex automations together into one bundled solution.
Creating an Agent#
Name & Description#
The name and description used only to identify your agents and make sharing with others easier.
These values are not used anywhere in the prompts or in the Agents instructions.
Instructions#
Instructions can impact an Agents decisions. Here you can give it extra guidance on how to approach using automations, what tone to use, how to respond etc.
If your automations are similar to eachother or if there is any nuance in when to use a certain automation over another, add extra instruction for better performance.
Picking Automations#
The Agents make a real-time decision on which automation to use during back and forth conversation.
Agents infer which automation to use based on the automation name, description and input names. This is very important because the quality of your automation names/decription could impact your agents decisions.
Naming an automation 'do the thing' with no description might cause some strange behaviour. A clear concise name + a detailed description will drastically improve how reliably your agent uses the correct automation.
Building an Agent Friendly Automation#
Any AgentHub automation can be run via an Agent but here are some important things to note to optimize the experience.
Inputs#
Input nodes are needed for Agents to provide input. If you want your agent to be able to decide on dynamic inputs for your automation you will need input nodes.
Lets take the example of sending an email. If you want your Agent to be able to decide on the recipient, the subject and the body you will need 3 input nodes feeding into a send email node.
The 'input_name' associated with that input node is what the agent will use to understand what the field should be populated with. In this email example if you named all three input nodes 'input', the agent would not know where to put the recipient address.
Outputs#
Output nodes are essential for the flow of conversation. The content of your output nodes will be passed to your agent once finished so it can formualate a message. If your agent uses an automation without an output node it tends to not know what to say when the run is complete and can lead to hallucinations.
Any completion information you can pass in the output is good. A common pattern is to pass the end product link into the output so the agent can formulate a nice completion message and serve you that link effectively.
What the Agent Can See#
Agents do not know what goes on within the automation. All they can see are the name, description and inputs. Automations can be infinitely complex but as long as it's clear when to use it, the agent will have no problem.
Sharing Agents#
To share an Agent with others copy the agent URL or from the hub select 'copy link' on an individual agent.
The conversation history is tied to your account. Revisiting an old Agent you've spoken to will reload your last conversation. Sending someone a link to your Agent will create a new conversation history for them. No one else can see your conversation history with any Agent, even their own.
Agent File Upload#
We currently don't support files in chat but will be adding this as a feature very soon.