Laravel AI SDK v1.0 — What's New | Mohamed Said [Skip to content](#main) [ ![](https://cdn.msaied.com/01KT78WE565VEMM3PSNQAAB0MH.png) Mohamed SaidLaravel Backend Engineer ](https://msaied.com/public) - [Home](https://msaied.com/public) - [Projects](https://msaied.com/public/projects) - [Articles](https://msaied.com/public/articles) - [Certificates](https://msaied.com/public/certificates) - [About](https://msaied.com/public#about) [ Contact](https://msaied.com/public#contact) Menu Menu ---- Close - [HomeStart here](https://msaied.com/public) - [ProjectsCase studies](https://msaied.com/public/projects) - [ArticlesEngineering notes](https://msaied.com/public/articles) - [CertificatesCredentials](https://msaied.com/public/certificates) - [AboutHow I work](https://msaied.com/public#about) - [ContactGet in touch](https://msaied.com/public#contact) [Start a conversation](https://msaied.com/public#contact) [WhatsApp](https://wa.me/201094619204) [Email](mailto:hello@msaied.com) 1. [Home](https://msaied.com/public) 2. / 3. [Articles](https://msaied.com/public/articles) 4. / 5. [Laravel](https://msaied.com/public/articles?category=laravel) 6. / 7. Laravel AI SDK v1.0: Classification, Approvable Tools, Agent Middleware & More [Laravel](https://msaied.com/public/articles?category=laravel) [AI](https://msaied.com/public/articles?category=ai) Laravel AI SDK v1.0: Classification, Approvable Tools, Agent Middleware & More =================================================================================== Laravel AI SDK v1.0 ships with Jev classification, approvable tool calls, per-step agent middleware, Vercel Chat and AG-UI protocol support, tool search, code execution, and a redesigned conversation schema. ![](https://cdn.msaied.com/01M22N44A70A5MC2S599JP0MPH.webp) [Mohamed Said](https://msaied.com/public#person) Published 23 Sep 2026 · Updated 24 Sep 2026 · 5 min read ShareCopy linkCopied ![Laravel AI SDK v1.0: Classification, Approvable Tools, Agent Middleware & More](https://cdn.msaied.com/698/65d5ceeb67b5ceb383a05b0d760515d1.png) On this page +1. [Laravel AI SDK v1.0 Is Here](#laravel-ai-sdk-v10-is-here) 2. [Jev Classification: A New Kind of AI Capability](#jev-classification-a-new-kind-of-ai-capability) 3. [Vercel Chat and AG-UI Protocol Support](#vercel-chat-and-ag-ui-protocol-support) 4. [Per-Step Agent Middleware](#per-step-agent-middleware) 5. [Approvable Tool Calls](#approvable-tool-calls) 6. [Redesigned Conversation Storage](#redesigned-conversation-storage) 7. [Tool Search and Code Execution](#tool-search-and-code-execution) 8. [Unified Token Usage Reporting](#unified-token-usage-reporting) 9. [Upgrading to v1.0](#upgrading-to-v10) 10. [Key Takeaways](#key-takeaways) Laravel AI SDK v1.0 Is Here --------------------------- Released on 23 September 2026, Laravel AI SDK v1.0 graduates the package from beta to a production-ready, Laravel-native interface for every major AI provider. The release closes hundreds of bugs and ships several headline features worth knowing before you upgrade. --- Jev Classification: A New Kind of AI Capability ----------------------------------------------- Not every AI task is text generation. Routing a support ticket, flagging a comment, or branching a workflow are *classification* decisions — and they should be fast and cheap. The SDK now treats classification as a first-class capability alongside text, images, audio, and embeddings. Jev (by TypeSafe) answers classification questions in milliseconds at a fraction of the cost of a traditional LLM. For yes/no questions, a new `decide` macro on the `Str` class returns a boolean. A `threshold` argument controls the required confidence level before the model commits to an answer. Classification currently runs on TypeSafe and OpenRouter, with more providers planned. --- Vercel Chat and AG-UI Protocol Support -------------------------------------- The SDK now speaks two popular frontend streaming protocols out of the box: - **Vercel Chat** — pass the incoming request to `Vercel::chat` and pipe the result straight to your agent. The single route handles message ingestion, conversation persistence, and tool-approval decisions. - **AG-UI** — used by clients such as CopilotKit, letting you stream agent responses to rich UI components without custom glue code. A helper method converts stored conversation messages back into the shape the client expects, making page-reload recovery straightforward. --- Per-Step Agent Middleware ------------------------- Previously, middleware ran once per prompt. It now wraps **each generation step**, so a three-tool agent run triggers your middleware three times. Each step arrives as a `PendingStep` you can inspect and mutate. Practical uses include: - Removing an expensive tool after it has already been called once - Swapping the model mid-prompt based on accumulated cost - Summarising a long message history before the next step - Returning a cached answer without hitting the provider at all Methods like `withModel`, `withTools`, `withMessages`, and `withMaxTokens` let you rewrite every part of the prompt between steps. --- Approvable Tool Calls --------------------- Some tools must not run without a human in the loop. Implement the `Approvable` contract on a tool and use the `InteractsWithApprovals` trait. When the agent reaches that tool, it pauses and returns a response describing what is waiting — including the arguments the model chose. Resume the conversation by submitting a decision for each pending call: **approve**, **reject** (with a reason fed back to the model), or **edit** the arguments before the tool runs. Approvals work with `prompt`, `stream`, `queue`, and broadcast methods. --- Redesigned Conversation Storage ------------------------------- The `conversations` table now stores a `steps` JSON column with one entry per round-trip, and each tool result is stored alongside the call that produced it. This eliminates the guessing required when replaying history for providers and makes approval state unambiguous. Pagination and the `toolCalls` / `toolResults` accessors remain unchanged. Raw SQL against the old `tool_calls` or `tool_results` columns must migrate to `steps`. A backfill migration is included in the upgrade guide and must be run before deploying v1.0. --- Tool Search and Code Execution ------------------------------ **Tool Search** wraps rarely used tools in a `ToolSearch` container so they are only loaded when the prompt actually calls for them — reducing token cost and improving model accuracy on large tool sets. Supported on OpenAI and Anthropic. **Code Execution** is a new provider tool that runs code in the provider's own sandbox, useful for data analysis and calculations. Supported on Anthropic, OpenAI, Azure, Gemini, and xAI. --- Unified Token Usage Reporting ----------------------------- The `promptTokens` and `completionTokens` properties are renamed to `inputTokens` and `outputTokens` across every provider. Cached, cache-written, and reasoning tokens are now reported as subsets of those totals rather than separate fields. --- Upgrading to v1.0 ----------------- Breaking changes touch conversation storage, agent middleware, token usage property names, and stream protocols. The [official upgrade guide](https://github.com/laravel/ai/blob/1.x/UPGRADE.md) covers each change with before/after examples and an impact rating. For AI-assisted upgrades, install **Laravel Boost** (the first-party MCP server) as a dev dependency and run the `/upgrade-ai-sdk-v1` slash command in Claude Code, Cursor, OpenCode, Gemini, or VS Code. --- Key Takeaways ------------- - **Jev classification** adds a fast, cheap, boolean-friendly AI capability to the SDK - **Vercel Chat and AG-UI** protocol support removes frontend integration boilerplate - **Per-step middleware** gives fine-grained control over cost, context, and model choice - **Approvable tools** let humans review and edit AI decisions before execution - **`steps` column** makes conversation replay reliable across all providers - **Tool Search** cuts token usage for agents with large tool sets - **Code Execution** brings sandboxed computation to five major providers - Token usage properties are renamed; update any monitoring or billing code --- Read the full announcement and documentation at . - [Laravel](https://msaied.com/public/articles?search=Laravel) - [AI SDK](https://msaied.com/public/articles?search=AI%20SDK) - [PHP](https://msaied.com/public/articles?search=PHP) - [Agents](https://msaied.com/public/articles?search=Agents) - [LLM](https://msaied.com/public/articles?search=LLM) - [Classification](https://msaied.com/public/articles?search=Classification) Frequently asked questions --------------------------- What breaking changes does Laravel AI SDK v1.0 introduce?The main breaking changes are: the conversations table now uses a `steps` JSON column instead of flat tool\_call and tool\_result lists (requiring a backfill migration); agent middleware now runs per step rather than per prompt; token usage properties are renamed from `promptTokens`/`completionTokens` to `inputTokens`/`outputTokens`; and stream protocol handling has changed. The official upgrade guide covers each change with before/after code examples. How do approvable tool calls work in Laravel AI SDK v1.0?Implement the `Approvable` contract on any tool and add the `InteractsWithApprovals` trait. When the agent reaches that tool during a run, it pauses and returns a response listing what is waiting for approval, including the arguments the model selected. You resume the conversation by submitting a decision — approve, reject with a reason, or edit the arguments — before the tool executes. This works with prompt, stream, queue, and broadcast methods. What is Jev classification and when should I use it instead of a standard LLM?Jev is a System One classification model from TypeSafe that answers classification questions in milliseconds at a fraction of the cost of a traditional LLM. Use it for decisions that don't require text generation — routing a ticket, flagging content, or branching a workflow. The SDK exposes it as a first-class capability and adds a `Str::decide()` macro for simple yes/no questions with a configurable confidence threshold. ![Mohamed Said](https://cdn.msaied.com/01M22N44A70A5MC2S599JP0MPH.webp)About the author ---------------- [Mohamed Said](https://msaied.com/public#person)Senior Backend Engineer specializing in Laravel, scalable SaaS platforms, APIs, and cloud infrastructure. I build secure, high-performance web applications that help businesses grow. [About](https://msaied.com/public#about) [GitHub ↗](https://github.com/EG-Mohamed) [LinkedIn ↗](https://www.linkedin.com/in/msaiedm/) [WhatsApp ↗](https://wa.me/201094619204) [Email Address ↗](mailto:hello@msaied.com) [My CV ↗](https://drive.google.com/file/u/0/d/1MF20IPRJyzfy32mhEutjL5EpSls0w2Q8/view) [Previous articleWhat's New in Laravel 13.33: Tagged Memoized Cache, Model Refreshes, and More](https://msaied.com/public/articles/whats-new-in-laravel-1333-tagged-memoized-cache-model-refreshes-and-more) [Next articleLaravel AI SDK 1.0: Classification, Tool Approvals, and Vercel Chat Streaming](https://msaied.com/public/articles/laravel-ai-sdk-10-classification-tool-approvals-and-vercel-chat-streaming) On this page ------------- 1. [Laravel AI SDK v1.0 Is Here](#laravel-ai-sdk-v10-is-here) 2. [Jev Classification: A New Kind of AI Capability](#jev-classification-a-new-kind-of-ai-capability) 3. [Vercel Chat and AG-UI Protocol Support](#vercel-chat-and-ag-ui-protocol-support) 4. [Per-Step Agent Middleware](#per-step-agent-middleware) 5. [Approvable Tool Calls](#approvable-tool-calls) 6. [Redesigned Conversation Storage](#redesigned-conversation-storage) 7. [Tool Search and Code Execution](#tool-search-and-code-execution) 8. [Unified Token Usage Reporting](#unified-token-usage-reporting) 9. [Upgrading to v1.0](#upgrading-to-v10) 10. [Key Takeaways](#key-takeaways) ### Have a technical challenge? Tell me what you’re building. I reply within two working days. [Start a conversation](https://msaied.com/public#contact) Related articles ----------------- [ ![](https://cdn.msaied.com/745/8744e1be5136b430da52e9fca3ed3964.png) · 3 min read### Service Container Deep Dive: Contextual Binding, Tagging, and Method Injection 6 Oct 2026 ](https://msaied.com/public/articles/service-container-deep-dive-contextual-binding-tagging-and-method-injection-1) [ ![](https://cdn.msaied.com/743/8998fac3a41451ab3fe1588194e17a43.png) Filament · 3 min read### Securing Filament Plugins with Plumb: Automated Security Scoring for PHP Packages 5 Oct 2026 ](https://msaied.com/public/articles/securing-filament-plugins-with-plumb-automated-security-scoring-for-php-packages) [ ![](https://cdn.msaied.com/742/2d02018669cdeedccb5de2efb898f0ee.png) Filament · 3 min read### Filament v3.3.56 Released: File Hash Names and Livewire Upload Fix 5 Oct 2026 ](https://msaied.com/public/articles/filament-v3356-released-file-hash-names-and-livewire-upload-fix) Have a technical challenge? ---------------------------- Tell me what you’re building. I reply within two working days. [Discuss your project ↗](https://msaied.com/public#contact) © 2026 Mohamed Said · Built with Laravel, meant to last.Senior Backend Engineer specializing in Laravel, scalable SaaS platforms, APIs, and cloud infrastructure. I build secure, high-performance web applications that help businesses grow. - [Home](https://msaied.com/public) - [Articles](https://msaied.com/public/articles) - [Certificates](https://msaied.com/public/certificates) - [GitHub](https://github.com/EG-Mohamed) - [LinkedIn](https://www.linkedin.com/in/msaiedm/) - [WhatsApp](https://wa.me/201094619204) - [Email Address](mailto:hello@msaied.com) - [My CV](https://drive.google.com/file/u/0/d/1MF20IPRJyzfy32mhEutjL5EpSls0w2Q8/view) - [Sitemap](https://msaied.com/public/sitemap.xml)