Laravel AI SDK 1.1 Adds Agent Skills and Cohere Support

The open-source laravel/ai package, which provides a unified API for querying OpenAI, Anthropic, Gemini and other providers from a Laravel application, reached version 1.1.0 on October 5, 2026. The update adds support for the Agent Skills format and brings Cohere into the list of text-generation providers. Two days later, a 1.2.0 release fixed several structured-output behaviors, a sign of the steady release pace since version 1.0 shipped in late September.

#A minor version, two additions that matter

Announced on September 23, 2026, the SDK’s 1.0 release laid the groundwork: multi-provider model calls, classification, and human approval (tool approval) on sensitive tools. Version 1.1.0 pushes the PHP ecosystem a step further on two fronts: adopting a skills format shared with other agent environments, and opening up to a seventh text provider.

#Agent Skills, now supported natively

A skill is structured as a folder containing a SKILL.md file, whose frontmatter must include a description; the skill’s name is optional and defaults to the folder name. An agent implementing the new HasSkills contract automatically gets a LoadSkill tool that lists available skills, which can come from directories, PHP Skill objects, or closures loaded on demand. Files over 256 KB and binary files are rejected by the loader.

The choice brings Laravel closer to a format already used in other agent environments, a topic Mail Studio covered in its pieces on the MCP specification and on editorial agents in production.

#Cohere joins the list of text providers

Cohere joins the list of supported text-generation providers via its v2 Chat API: streaming, tool calls, structured output and reasoning are all covered, with command-a-03-2025 as the default model. Document attachments and citations aren’t supported yet. The release also lets teams require human approval on the similarity-search and filesystem tools, alongside the MCP and agent tools already covered — an option that is off by default.

#Other changes, and already a 1.2.0

Version 1.1.0 also fixes provider failover: a tool unsupported by the fallback provider is now skipped instead of stopping the run. It also adds self-hosted classification via Laya and more tool mappings for Azure, Groq and OpenRouter. Version 1.2.0, released on October 7, in turn fixes structured output on Bedrock and updates Anthropic’s default models.

#What it means for PHP teams

By aligning its skills format with a shared standard rather than a proprietary convention, the SDK makes it easier to reuse agent instructions written for other environments. That positions Laravel as a credible base for agent backends, alongside the Node and Python ecosystems more often cited on this ground.

class SupportAgent implements HasSkills
{
    public function skills(): array
    {
        return [
            // Folder containing resources/skills/refund-policy/SKILL.md
            SkillDirectory::make(resource_path('skills/refund-policy')),

            // Skill declared directly in PHP
            Skill::make('escalation')
                ->description('Decides when to hand off a ticket to a human.'),

            // Lazy-loaded, e.g. from the database
            fn () => Skill::fromDatabase('product-faq'),
        ];
    }
}

The model only loads a skill’s full instructions when a task calls for it, through a single LoadSkill tool exposed to the agent.

#Key takeaways

• Version 1.1.0 of laravel/ai, released on October 5, 2026, adds the Agent Skills format and the Cohere provider.
• Skills load on demand through the HasSkills contract and the LoadSkill tool, with a 256 KB per-file limit.
• Human approval on sensitive tools stays optional and must be turned on explicitly.
• A 1.2.0 release, two days later, already fixed several structured-output issues.

Further reading: the detailed announcement on Laravel News and the release notes on GitHub.