Neuron Streaming Step

This example shows a clean split:

Neuron streams the agent response chunk by chunk
Queuety persists each chunk durably through StreamingStep

That means the UI can receive a live draft, while the workflow still survives worker crashes or retries.

When this pattern is useful

Use it when:

a draft or response is long enough that you want to stream it live
you do not want to lose partial output if the worker dies
the final artifact should still land in workflow state after the stream completes

The Neuron agent

namespace App\Neuron;

use NeuronAI\Agent\Agent;
use NeuronAI\Providers\AIProviderInterface;
use NeuronAI\Providers\OpenAI\Responses\OpenAIResponses;
use NeuronAI\SystemPrompt;

final class DraftMemoAgent extends Agent
{
    protected function provider(): AIProviderInterface
    {
        return new OpenAIResponses(
            key: $_ENV['OPENAI_API_KEY'],
            model: $_ENV['OPENAI_MODEL'],
        );
    }

    public function instructions(): string
    {
        return (string) new SystemPrompt(
            background: [
                'You write crisp internal strategy memos.',
                'Prefer short sections and operationally useful language.',
            ],
        );
    }
}

The Queuety streaming step

namespace App\Workflow\Steps;

use App\Neuron\DraftMemoAgent;
use NeuronAI\Chat\Messages\UserMessage;
use Queuety\Contracts\StreamingStep;
use Queuety\Heartbeat;

final class StreamMemoDraftStep implements StreamingStep
{
    public function stream(array $state, array $existing_chunks = []): \Generator
    {
        if (! empty($existing_chunks)) {
            return;
        }

        $handler = (new DraftMemoAgent())->stream(
            new UserMessage(sprintf(
                "Write a strategy memo about %s for account %d.",
                $state['topic'],
                $state['account_id'],
            ))
        );

        foreach ($handler->events() as $i => $chunk) {
            Heartbeat::beat(['chunks' => $i + 1]);

            // This example assumes the agent has no tools attached, so
            // Neuron will stream text/reasoning chunks only.
            yield $chunk->content;
        }
    }

    public function on_complete(array $chunks, array $state): array
    {
        return [
            'draft_markdown' => implode('', $chunks),
        ];
    }

    public function config(): array
    {
        return [
            'max_attempts' => 3,
        ];
    }
}

Use it in a workflow

use Queuety\Queuety;

Queuety::workflow('memo_drafting')
    ->then(BuildDraftBriefStep::class)
    ->then(StreamMemoDraftStep::class)
    ->then(StoreMemoStep::class)
    ->dispatch([
        'account_id' => 10,
        'topic' => 'pricing changes for enterprise accounts',
    ]);

Why use Queuety instead of Neuron stream adapters alone

Neuron's stream adapters are useful for frontend protocols and transports.

Queuety adds something different:

every yielded chunk is persisted
retries can resume with the previously saved chunks
the completed draft is merged back into workflow state
the stream becomes one durable step in a larger orchestration

That makes this pattern a good fit when the stream is part of a multi-step business process, not just a request/response UI.

If your Neuron agent uses tools while streaming, handle the streamed chunk types explicitly. Neuron's current streaming docs describe TextChunk, ReasoningChunk, ToolCallChunk, and ToolResultChunk. For a background drafting workflow, the simplest approach is usually to keep the streaming agent text-only.