Playground

The Agent Playground is an interactive sandbox where you can chat with any discovered agent in real time. Test prompts, debug behavior, and experiment with parameters — all without writing a single line of code.

Agent List

Access at /ai-orbit/playground. The playground lists all agents discovered by scanning your configured directories.

Discovery

Orbit scans the directories configured in config/ai-orbit.php:

'agent_directories' => [
    'app/AI/Agents',
    'app/Ai/Agents',
],

Agents are discovered using PHP's tokenizer — no file loading or class instantiation happens during scanning, making discovery fast and safe.

Agent Cards

Each discovered agent is displayed as a card showing:

• Class Name — Fully qualified class name
• Tools — Count of registered tools (if implements HasTools)
• Structured Output — Badge indicating schema support
• Temperature — Default temperature (from attribute or method)

Click any agent card to enter the sandbox.

Agent Sandbox

Access at /ai-orbit/playground/{agent}. The sandbox provides a fully functional chat interface for the selected agent.

Intelligent Dependency Resolution

One of Orbit's most powerful features is automatic constructor dependency resolution. When you open a sandbox, Orbit analyzes the agent's constructor and classifies each parameter:

Strategy	What Happens	Example
Eloquent Picker	Dropdown populated with recent records	User $user, Order $order
Container	Auto-resolved from Laravel's service container	LoggerInterface $logger
Default	Uses the parameter's default value	int $limit = 10
Input	Text field for manual entry	string $apiKey
Unresolvable	Flagged as unavailable for sandbox	Complex custom types

Eloquent Model Picker

When an agent requires an Eloquent model, Orbit:

1. Detects the model class
2. Identifies display columns (name, title, subject, email, label, slug)
3. Fetches recent records (configurable, default 20)
4. Presents a searchable dropdown

class InvoiceAgent implements Agent
{
    public function __construct(
        public Invoice $invoice,  // Orbit shows a picker with recent invoices
        public User $user,         // Orbit shows a picker with recent users
    ) {}
}

Live Parameter Overrides

Override agent behavior on the fly without touching code:

Override	Description
System Prompt	Replace the agent's instructions entirely
Model	Switch to a different AI model
Provider	Switch to a different AI provider
Temperature	Adjust creativity/randomness (0–2)
Max Tokens	Limit response length

Changes take effect immediately on the next message.

Multi-Turn Chat

The sandbox supports persistent conversation sessions:

• Send multiple messages in a single session
• Conversation context is maintained automatically
• Tool calls and results are displayed inline
• Clear the session to start fresh

Tool Call Visualization

When an agent uses tools, the sandbox displays:

1. Tool Call — The tool name and arguments (expandable JSON)
2. Tool Result — The execution result
3. Final Response — The agent's response after tool execution

Session Persistence

Sandbox sessions are persisted in the session. If you refresh the page, your conversation and parameter inputs are restored.

Clear the session at any time with the Clear button.

How It Works

AgentIntrospector

The AgentIntrospector service analyzes agent constructors:

use Ashrafic\AiOrbit\Services\AgentIntrospector;

$introspector = app(AgentIntrospector::class);

// Analyze constructor
$analysis = $introspector->analyzeConstructor(InvoiceAgent::class);
// Returns: resolvable, needs_input, params[]

// Resolve parameters from user inputs
$resolved = $introspector->resolveParams(InvoiceAgent::class, [
    'invoice' => 42,
    'user' => 7,
]);

// Get recent records for a model
$records = $introspector->getModelRecords(Invoice::class, limit: 20);

SandboxRunner

The SandboxRunner executes the agent:

use Ashrafic\AiOrbit\Services\SandboxRunner;

$runner = app(SandboxRunner::class);

$result = $runner->execute(
    agentClass: InvoiceAgent::class,
    prompt: 'Summarize this invoice',
    paramInputs: ['invoice' => 42, 'user' => 7],
    overrides: ['model' => 'gpt-4-turbo', 'provider' => 'openai'],
);

$result->content;          // The agent's response
$result->mode;             // 'full' (real execution)
$result->sdkConversationId; // The conversation ID
$result->toolCalls;        // Array of tool calls
$result->toolResults;      // Array of tool results

Use Cases

Rapid Prototyping

Test new prompts and parameters against your agents without deploying code changes.

Debugging Production Issues

Reproduce a production conversation by selecting the same Eloquent model and sending similar prompts.

Parameter Tuning

Experiment with temperature and max tokens to find the optimal settings for your use case.

Model Comparison

Test the same prompt across different models using the live override feature.

Training Demo

Show stakeholders how your agents work by interacting with them in a clean, isolated sandbox.

Limitations

• Agents with unresolvable constructor dependencies cannot be simulated
• Complex nested dependencies may not be fully resolved
• The sandbox uses the real AI provider, so it consumes real API credits

Customization

Override the sandbox views:

php artisan vendor:publish --tag=ai-orbit-views

Then edit:

• resources/views/vendor/laravel-ai-orbit/playground/index.blade.php — Agent list
• resources/views/vendor/laravel-ai-orbit/playground/show.blade.php — Sandbox layout
• resources/views/vendor/laravel-ai-orbit/livewire/agent-sandbox.blade.php — Sandbox component
• resources/views/vendor/laravel-ai-orbit/livewire/agent-inspector.blade.php — Inspector panel