Experimenting multi agent applications on Prism

[grpaiva]

First of all, I love Prism. Thank you so much for all the hard work on this amazing package!

TL;DR

I built a small experimental package called prism-agents on top of Prism, inspired by OpenAI's AgentsSDK.
It introduces concepts like Agents, Agent Runners, Handoffs, Guardrails, and Tracing — all designed to feel as fluent and easy to use as native Prism.
It’s still a prototype, but already works for chaining agents, building agent tools, and tracing agent runs into a database.
Would love any feedback, ideas, or if anyone finds it useful!

Long version

I’m familiar with a few libraries for building multi-agent applications in Python, like LangChain, CrewAI, and now OpenAI’s AgentsSDK. However, I’ve felt for a while that something like this was missing in our ecosystem.

So, I built a small package for use in my own projects. It’s heavily inspired by OpenAI’s AgentsSDK and uses Prism under the hood. It implements a few concepts that I find pretty useful:

• Agents and Agent Runners
• Agent Handoffs
• Guardrails
• Tracing

I tried to design it to feel just like using Prism — using the same builder pattern and even some of the same method names — to provide a fluent and easy-to-use interface.

Usage examples:

Agent as tools

use Grpaiva\PrismAgents\Agent;
use Grpaiva\PrismAgents\PrismAgents;
use Prism\Prism\Enums\Provider;

// Create specialized translation agents
$spanishAgent = Agent::as('spanish_agent')
    ->withInstructions('You translate the user\'s message to Spanish')
    ->using(Provider::OpenAI, 'gpt-4.1-nano')
    ->withHandoffDescription('An english to spanish translator');

$frenchAgent = Agent::as('french_agent')
    ->withInstructions('You translate the user\'s message to French')
    ->using(Provider::OpenAI, 'gpt-4.1-nano')
    ->withHandoffDescription('An english to french translator');

// Create an orchestrator agent and add other agents as tools
$orchestratorAgent = Agent::as('orchestrator_agent')
    ->withInstructions('You are a translation agent. Use the tools given to you to translate.')
    ->using(Provider::OpenAI, 'gpt-4.1-mini')
    ->withMaxSteps(10)
    ->withTools([
        $spanishAgent->asTool(),
        $frenchAgent->asTool()
    ]);

// Run the orchestrator
$result = PrismAgents::run($orchestratorAgent, "Translate 'Hello, how are you?' to Spanish and French.")
    ->withTrace('translation_process');

// Output the result
echo $result->getOutput();

Guardrails

use Grpaiva\PrismAgents\Agent;
use Grpaiva\PrismAgents\PrismAgents;
use Grpaiva\PrismAgents\Guardrail;
use Grpaiva\PrismAgents\GuardrailResult;
use Grpaiva\PrismAgents\AgentContext;
use Prism\Prism\Enums\Provider;

// Create a custom guardrail
class ProfanityGuardrail extends Guardrail
{
    public function check($input, AgentContext $context): GuardrailResult
    {
        // Simple example - check for a banned word
        if (is_string($input) && stripos($input, 'badword') !== false) {
            return GuardrailResult::fail('Input contains profanity', 400);
        }
        
        return GuardrailResult::pass();
    }
}

// Create an agent with the guardrail
$agent = Agent::as('safe_assistant')
    ->withInstructions('You are a helpful assistant.')
    ->using(Provider::OpenAI, 'gpt-4o')
    ->withInputGuardrails([ProfanityGuardrail::as('profanity_filter')]);

try {
    // This will throw a GuardrailException if the input fails the guardrail
    $result = PrismAgents::run($agent, "Hello, can you help me with badword?");
    echo $result->getOutput();
} catch (\Grpaiva\PrismAgents\Exceptions\GuardrailException $e) {
    echo "Guardrail triggered: " . $e->getMessage();
}

Tracing

I find the tracing part particularly useful. Right now, you can choose a database driver, and traces will be stored in a prism_agent_traces table. For convenience (mainly for myself), I’ve also created some views, inspired by OpenAI’s and LangSmith’s tracing panels:

Traces table:

Trace view:

This project is still a prototype, although I’m already using it in a couple of personal projects. It’s far from production-ready. I also leveraged a lot of AI to help build it — it almost feels like it was created through "vibe coding." But I’m happy with how it's taking shape.

I’m not entirely sure why I’m sharing all this — maybe some concepts could eventually be implemented into Prism itself, or perhaps others might find it useful or want to contribute. Either way, I thought it would be worth sharing.

Cheers!

Edit 1: Added missing php language tags to code blocks for proper syntax highlighting.

Edit 2: Added Guardrails example and removed simple agent with tool example.

[emreakay]

thats realy cool

[CeriseGoutPelican]

This is absolutely awesome and super timely for the project I am working on! Will definitely keep an eye on the development of the package. Keep up the great work!

[mhphilip]

Great stuff. Would love to see most of this in the prism core. That would really make it the go-to solution for agentic php development.

[CeriseGoutPelican]

100%

[CeriseGoutPelican]

Hey @grpaiva, I tried to test your package earlier today, but it doesn’t seem to be available on Packagist.

And looking a bit more into your documentation, I agree with @mhphilip that it should be added to Prism core — especially considering how closely agents and tools are related logically, and with the recent opentelemetry release.

$ composer show grpaiva/prism-agents -a
Package "grpaiva/prism-agents" not found.  

[grpaiva]

Hey @CeriseGoutPelican, my bad! I just tagged and published it. I'm already using it in a couple of applications but haven't had much time to work on it lately. I see plenty of improvements to make, but it does get the job done for now. Cheers!

[CeriseGoutPelican]

Amazing, thanks a lot!

Very happy to contribute if I end up using it in the app I’m currently developing.

---

I'm currently testing a Deep Search agent using LLM reranking and Scout 👀