Abilenduke Copilot Developer AI Agent Patterns

AI Agent Patterns

When to Apply

• Creating a new agent class in app/Ai/Agents/
• Modifying AgentResolver action or prompt mappings
• Writing tests for agents (::fake(), ::assertPrompted())
• Working with SSE streaming in AdminArticleAiController
• Adding or changing apply block formats (:::apply-body, :::apply-excerpt)
• Debugging agent responses, conversation linking, or temperature behavior

Documentation

Use search-docs for detailed Laravel AI SDK documentation before making changes.

Agent Anatomy

Every agent implements Agent & Conversational (intersection type), uses three traits, and is configured via PHP attributes. All agents run on Ollama (Docker service), not OpenAI.

use App\Ai\Agents\Concerns\HasArticleContext; use Laravel\Ai\Attributes\MaxTokens; use Laravel\Ai\Attributes\Model; use Laravel\Ai\Attributes\Temperature; use Laravel\Ai\Attributes\Timeout; use Laravel\Ai\Concerns\RemembersConversations; use Laravel\Ai\Contracts\Agent; use Laravel\Ai\Contracts\Conversational; use Stringable;

#[Model('llama3.2:3b')] // Ollama model — NOT OpenAI #[Temperature(0.6)] // 0.3 (precise) to 0.8 (creative) #[MaxTokens(4096)] // Fixed for all agents #[Timeout(120)] // Fixed for all agents class ArticleRefiner implements Agent, Conversational { use HasArticleContext, Promptable, RemembersConversations;

public function instructions(): Stringable|string
{
    return <<<INSTRUCTIONS
    You are an expert content editor...

    {$this->articleContextBlock()}
    {$this->applyBlockInstructions()}
    INSTRUCTIONS;
}

}

Temperature spectrum — choose based on the agent's task precision:

Non-conversational agents: HumanizerAgent implements only Agent (NOT Conversational) because it processes text in a single pass without conversation memory. It does NOT use HasArticleContext — it has its own constructor signature.

Content Generation Patterns

The HasArticleContext trait provides the constructor, context accessors, and prompt-building methods shared by all conversational agents.

protected function contextTitle(): string { return $this->context['title'] ?? ''; }
protected function contextBody(): string { return $this->context['body'] ?? ''; }
protected function contextExcerpt(): string { return $this->context['excerpt'] ?? ''; }
protected function contextContentType(): string { return $this->context['content_type'] ?? ''; }
protected function contextCategory(): string { return $this->context['category'] ?? ''; }
protected function estimateWordCount(string $body): int { return str_word_count(strip_tags($body)); }

}

Prompt-building methods (called in instructions()):

• articleContextBlock() — Returns article metadata (title, type, category, word count) + writing quality standards (banned phrases from config('humanizer.banned_phrases'), sentence variety rules)
• applyBlockInstructions() — Returns formatting rules for apply blocks

Apply blocks tell the frontend to auto-apply content back to the article editor:

// For replacing article excerpt: :::apply-excerpt A compelling 1-2 sentence summary under 160 characters. :::

When to use apply blocks: Only when the agent generates or rewrites replacement content. For feedback, suggestions, and analysis, agents respond in plain markdown without apply blocks. This distinction is enforced in each agent's instructions() prompt.

AgentResolver Routing

AgentResolver is the single routing layer between HTTP endpoints and agent classes. All agent instantiation goes through it — never new Agent() directly.

protected const INLINE_AGENTS = [
    'proofread' => Proofreader::class,
    'grammar'   => GrammarChecker::class,
];

// Each constant also has matching prompt templates:
// ACTION_PROMPTS and INLINE_PROMPTS

}

Factory methods — all return Agent&Conversational intersection type:

• forAction(string $action, array $context) — Quick actions (proofread, grammar, generate, refine, seo, excerpt)
• forInlineReview(string $action, array $context) — Inline editor actions (proofread, grammar only)
• forChat(array $context) — Conversational chat (always ArticleRefiner)
• forSelectionChat(array $context) — Freeform instruction on selected text (always ArticleRefiner)

To add a new agent:

1. Create the class in app/Ai/Agents/ implementing Agent, Conversational
2. Use HasArticleContext, Promptable, RemembersConversations traits
3. Add to ACTION_AGENTS and/or INLINE_AGENTS in AgentResolver
4. Add a prompt template to ACTION_PROMPTS and/or INLINE_PROMPTS
5. Update validActions() or validInlineActions() if adding new action names

return new $class($context);

}

// Prompt includes body context appended to the template public static function actionPrompt(string $action, array $context): string { $prompt = self::ACTION_PROMPTS[$action];

if ($action === 'generate') {
    $prompt .= "\n\nArticle body to base generation on:\n".substr(strip_tags($context['body'] ?? ''), 0, 500);
} elseif (in_array($action, ['proofread', 'grammar', 'seo', 'refine'])) {
    $prompt .= "\n\nArticle body:\n".($context['body'] ?? '');
} elseif ($action === 'excerpt') {
    $prompt .= "\n\nArticle body:\n".substr(strip_tags($context['body'] ?? ''), 0, 2000);
}

return $prompt;

}

Streaming & Conversation

The controller uses two streaming methods depending on whether conversations should persist.

return response()->stream(function () use ($streamable) {
    try {
        foreach ($streamable as $event) {
            echo 'data: '.((string) $event)."\n\n";
            if (ob_get_level() > 0) { ob_flush(); }
            flush();
        }
    } catch (\Throwable $e) {
        echo 'data: '.json_encode(['type' => 'error', 'message' => $e->getMessage()])."\n\n";
        flush();
        report($e);
    }
    echo "data: [DONE]\n\n";
    flush();
}, 200, [
    'Content-Type' => 'text/event-stream',
    'X-Accel-Buffering' => 'no',
    'Cache-Control' => 'no-cache',
    'Connection' => 'keep-alive',
]);

}

Conversation linking (for chat, not inline actions):

• New conversation: $agent->forUser($user) → starts fresh
• Resume: $agent->continue($conversationId, as: $user) → resumes existing
• After streaming: $agent->currentConversation() → get conversation ID
• Link to article: Update agent_conversations.article_id via .then() callback on the streamable
• Conversations stored in agent_conversations + agent_conversation_messages tables (UUID PKs)

Generate action is special: Dispatches ProcessArticle job instead of streaming. Returns 202 Accepted with JSON. The job runs generation server-side, humanizes via TextHumanizer, and broadcasts progress via Reverb.

$streamable = $agent->stream($prompt); $articleId = $validated['article_id'] ?? null;

$streamable->then(function () use ($agent, $articleId) { $conversationId = $agent->currentConversation();

if ($conversationId && $articleId) {
    DB::table('agent_conversations')
        ->where('id', $conversationId)
        ->whereNull('article_id')
        ->update(['article_id' => $articleId]);
}

});

Testing

Each agent class has its own ::fake() and ::assertPrompted() static methods provided by the Laravel AI SDK. Always fake the specific agent class, not a generic mock.

// Helper function for valid context (reuse across tests) function validContext(): array { return [ 'title' => 'Test Article', 'body' => 'Test article body with enough content.', 'excerpt' => 'A test excerpt', 'content_type' => 'general', 'category' => 'Technology', ]; }

// Fake an agent with a canned response test('chat endpoint streams response', function () { $admin = User::factory()->admin()->create(); ArticleRefiner::fake(['Improved content here']);

$this->actingAs($admin)
    ->postJson('/admin/articles/ai/chat', [
        'message' => 'Improve this article',
        'context' => validContext(),
    ])->assertOk();

ArticleRefiner::assertPrompted('Improve this article');

});

// Fake an agent for quick actions test('action endpoint resolves correct agent', function () { $admin = User::factory()->admin()->create(); Proofreader::fake(['Feedback here']);

$this->actingAs($admin)
    ->postJson('/admin/articles/ai/action', [
        'action' => 'proofread',
        'context' => validContext(),
    ])->assertOk();

Proofreader::assertPrompted(fn ($prompt) =>
    str_contains($prompt->prompt, 'Review the current article')
);

});

// Generate action uses Queue::fake() instead test('generate action dispatches job', function () { Queue::fake(); $admin = User::factory()->admin()->create(); $article = Article::factory()->create();

$this->actingAs($admin)
    ->postJson('/admin/articles/ai/action', [
        'action' => 'generate',
        'article_id' => $article->id,
        'context' => validContext(),
    ])->assertAccepted();

Queue::assertPushed(ProcessArticle::class);

});

Key testing patterns:

• AgentName::fake(['response text']) — Synchronous fake with canned response
• AgentName::fake() — Empty response fake (for validation tests)
• AgentName::assertPrompted('substring') — Assert the prompt contained text
• AgentName::assertPrompted(fn ($prompt) => ...) — Assert with closure for complex matching
• Queue::fake() + Queue::assertPushed(ProcessArticle::class) — For the generate action
• $this->withoutMiddleware(ValidateCsrfToken::class) — For JSON API tests
• Conversation history tests must seed agent_conversations and agent_conversation_messages tables directly with all required JSON columns (attachments, tool_calls as empty arrays/objects)

Common Pitfalls

• Thinking-capable models crash the SDK: Avoid models like qwen3 that emit ThinkingCompleteEvent — the Laravel AI SDK doesn't handle it and throws an unrecoverable exception. Stick to llama3.2:3b or similar non-thinking models.
• Ollama, not OpenAI: Agents use #[Model('llama3.2:3b')] which runs on the Ollama Docker service. Don't confuse this with the global config('ai.default_provider') which is for different use cases. The model attribute is the Ollama model name.
• Temperature mismatches: Using high temperature (0.8) for grammar checking produces inconsistent corrections. Using low temperature (0.3) for content generation produces flat, repetitive output. Match temperature to task precision — see the spectrum table above.
• Missing context fields cause silent failures: The context array has 5 fields (title, body, excerpt, content_type, category). The body field uses ['present', 'nullable'] validation — it must be present in the request even if null. Controllers coerce null to empty string: $validated['context']['body'] = $validated['context']['body'] ?? ''.
• Forgetting to register in AgentResolver: Creating a new agent class without adding it to ACTION_AGENTS or INLINE_AGENTS in AgentResolver means it's unreachable from the controller. Also add a prompt template to ACTION_PROMPTS/INLINE_PROMPTS and update validActions()/validInlineActions().

Source: ABilenduke/copilot-developer — distributed by TomeVault.