I can’t help but continue to dream how powerful AI Agents & Workflows could level up Fibery 1,000x. In hope of some community feedback or further collaboration, I’m sharing some initial progress I’ve made while tinkering with implementing an Agentic Workflow workspace in Fibery, albeit with all of its limitations. I’ve published this public Fibery whiteboard with my current thought process and concept progress.
KEY CONCEPTS
Workflow as an entity
Agentic Workflow entity represents a collection of inputs, steps, & outputs
Whiteboard linked to entity to visually build the workflow.
Input Blocks in Whiteboard
Elements on the whiteboard that may be used as “inputs” into downstream elements.
Would be nice to consider “Input Block” field types on the entity (or at least expose current entity fields as input options that can be displayed independently on the whiteboard).
AI Blocks in Whiteboard
Elements on the whiteboard that can receive inputs, a prompt, run an LLM command, and return a result.
Blocks could be an AI Action (direct call to an LLM) or AI Agent entity (see below).
AI Agent Entities as AI Blocks
As illustrated in the diagram above, I have an AI Agents database with fields: System prompt and User prompt (the user prompt accepts an input from a content block or previous workflow step).
Would be nice to consider “AI Agent” field type (or relation) on the entity. Given the user has added “AI Agents” to the entity in the entity page/form, when they navigate to the Whiteboard, those AI block elements will automatically be added to the whiteboard.
Historical Tracking of Workflow Runs & Agent Responses
I’ve created Agent Responses database, which tracks historical “runs” of an AI Agent entity. It’s basically inherits the AI Agent class, copying the System & User prompt at that point in time and the resulting “Response” for that agent run. This will help with viewing AI Agent efficacy over time and any improvements/degradations based on varying system/user prompts.
I’ve also created a concept “Workflow Runs” database to do the same, track historical completions of an Agentic Workflow run. This will help with viewing changes to the workflow over time–inputs, AI Agents/Actions, etc.
That’s it for now. I’ll be continuing this concept to see how far I can get. Hope this sparks some good conversation or interest in collborating to come up with a nifty Agent solution in Fibery.