RFC / architectural convergence: unified prompt rendering, lazy execution, and typed schemas across Rails AI frameworks

[Alexander-Senko]

Hi everyone,

Over the last year, I’ve been developing Action AI, an open-source framework exploring an Action Mailer / Action Controller mental model for LLM interactions in Rails.

As Active Agent grows and gains mindshare, it's clear that both libraries share a fundamental goal: preventing raw LLM provider calls from cluttering Rails controllers, jobs, and models by establishing a dedicated, maintainable interaction layer.

However, we are currently facing a classic open-source challenge: duplicating core responsibilities and controller dispatch abstractions. Both libraries implement their own controller/agent wrappers, action dispatches, and provider bindings. Having two distinct framework layers doing ~80% of the same structural routing runs the risk of fragmenting developer efforts across the Rails ecosystem.

Rather than running parallel tracks or forcing users into an either/or choice, I’d like to open a discussion on whether we can combine our efforts and converge on a unified standard for Rails AI interactions.

Key architectural strengths & innovations from Action AI

In Action AI, I’ve focused heavily on three specific capabilities that complement the broader AI ecosystem:

1. View-backed prompt engineering (ERB templates):
   Instead of defining prompts as Ruby strings or heavy configuration blocks, Action AI routes actions directly to standard Action View templates in app/ai/prompts/*.erb. Instance variables initialized in agent actions (@task, @user) are passed directly into the ERB context. This keeps complex multi-line prompt formatting, partials, and view helpers cleanly separated from application logic.

2. Lazy execution & chainable proxies:
   Calling an action method returns a lazy execution proxy. The prompt execution is deferred until #content, #parsed, #object, or #run is explicitly invoked. This makes it straightforward to chain actions or apply middleware-style modifications prior to dispatching:

WorkflowAgent.collect("release notes").refine(style: "bullet list").content

3. Active Model schema mapping:
   Action AI includes native type-mapping utilities (returns Person or returns [Person]) that automatically generate JSON schema contracts from Active Model attributes and wrap the returned structured output back into typed model instances.

The challenge: avoiding redundant controller layers

Because both Active Agent and Action AI treat the entry point as a controller/agent class, trying to turn Action AI into an external plugin for Active Agent creates an unnecessary double-wrapping problem. If Action AI loses its action routing, it gets stripped down to a plain set of small features like template renderer and schema wrapper. Conversely, wrapping Active Agent underneath Action AI creates two competing DSLs in the same project.

Proposal for discussion: How do we combine forces?

Instead of maintaining separate framework dispatches, how can we best bring these patterns together into Active Agent?

1. First-class view/template backend:
   Can Active Agent natively support Action AI's ActionView-backed template resolution (app/ai/prompts/...) so prompts can be written as clean templates rather than inline strings?

2. Standardized schema contracts:
   Can we unify structured output parsing so Active Model schemas, JSON Schema generation, and typed object instantiation are standardized across the gem?

3. Execution pipeline:
   Can we integrate lazy execution proxies to allow method chaining before triggering provider calls?

I am open to contributing these patterns, PRs, and architectural ideas directly into Active Agent so we can build a single, idiomatic, and robust AI standard for the Ruby on Rails community.

Would love to hear your thoughts on how we can structure this collaboration!

[sirwolfgang]

Have you reviewed ActiveAgents and the docs?

The current version of the gem supports these features

• https://docs.activeagents.ai/agents/instructions#format-support
• https://docs.activeagents.ai/agents/generation
• https://docs.activeagents.ai/actions/structured_output#from-activemodel

[Alexander-Senko]

That’s not literally what I meant, but I get the point. Thank you for pointing me out.