Laravel AI Development ServicesAgents, tools and document search inside your application

$response = (new TicketTriage)->prompt(
    $ticket->body,
    provider: [Lab::Anthropic, Lab::OpenAI],
    timeout: 120,
);

$ticket->update(['priority' => $response['priority']]);

Agents

An agent is a PHP class with its own instructions, created with php artisan make:agent. It can be prompted directly, streamed or queued.

Tools your application controls

php artisan make:tool defines what an agent may do in your system, with a schema for its arguments. Tools can also come from MCP servers.

Human approval

A tool that implements Approvable waits for a person before it runs, and can decide case by case. Deleting a record or sending a refund should work this way.

Structured output

An agent with HasStructuredOutput returns data matching a JSON schema, so your code can read a score or a category rather than parse prose.

Document search

Embeddings, reranking, vector stores and the SimilaritySearch tool let an agent answer from your own documents and data.

Conversations

The package's migrations create agent_conversations and agent_conversation_messages tables, so chat history is stored in your own database.

Streaming and queues

Stream replies over server-sent events, including the Vercel AI SDK protocol, or use queue() with then and catch callbacks for long-running work.

Failover

Pass a list of providers and the SDK moves on when one is rate-limited, overloaded or out of credits. A bad request does not trigger failover, which is what you want.

Images, audio and transcription

The same package covers image generation, text to speech and speech to text, each with its own list of supported providers.

Testing without API calls

Agents can be faked in tests with fixed responses, and assertPrompted checks what was sent, so the suite stays fast and costs nothing to run.

Source: Laravel AI SDK documentation.

See the Laravel upgrade service and its fixed prices

1. 01

   Pick one workflow

   We choose a task with a visible cost today, and agree how to judge the result, using a set of real examples from your data.

2. 02

   Prove it on your examples

   A small agent runs against those examples and we compare providers and models. If it does not do well enough, we tell you before you spend more.

3. 03

   Build it into the application

   Tools with least-privilege access, approval on anything that changes data, queues for slow calls, and tests that fake the provider.

4. 04

   Release and watch

   A staged release with failover configured, logging of prompts and outcomes as your data rules allow, and a review of cost and quality with you.