Laravel AI SDK 1.1 added Cohere, Agent Skills and approval gates for built-in tools

The post is a one-author walkthrough, not release notes. Each feature it details sits under a 1.1 heading, while its introduction groups 1.1 and 1.2 together ^[10]. The contents of 1.2 itself, classification included, are therefore unconfirmed. The package changelog is the place to settle it.

The post proposes three tiers for agent actions. Read-only operations retrieve information. Low-risk operations can run automatically under controlled conditions. Sensitive operations need authorization or explicit approval ^[12]. In our view that is the right default when an agent holds tools that can write.

The post states the limit itself. Tool approval is one part of a larger security strategy, and authentication, authorization, input validation, audit logs and server-side permission checks are still essential ^[9]. An approval step inside the agent loop does nothing for a code path that reaches the same action another way. The check belongs where the action executes. Whether the SDK's approval step pauses execution or only records the request is the first thing to test.

The post's example for skills is an ecommerce agent that needs one set of instructions for order lookup and another for product descriptions ^[11]. "Compatible agents" is the qualifier to check in the docs, because a skill only helps if the agent in use can load it ^[5].

The SDK's pitch is one interface across providers ^[2]. The post's own caveat limits that: provider compatibility is not universal, and supported features and tools vary by model and integration ^[7]. Cohere's confirmed scope is text generation ^[6]. A team that wants tool calls or structured output on Cohere should test each one.

Trying the package costs one Composer command, composer require laravel/ai, followed by provider credentials set up per the official documentation ^[3].