Agent Plugins 1.0.0 + Cloudflare AI Search Updates

Baseten joins Hugging Face Hub inference providers

Route LLM inference through Baseten directly from HF SDKs using OpenAI-compatible client calls; billing passes through provider costs with no markup.

Developers can now call Baseten-hosted models (DeepSeek V4 Flash, GLM-5.2, Kimi K3) via standard HF authentication without provider setup, reducing integration friction. Billing flexibility—either direct to provider or routed through HF account—lets teams choose cost/control tradeoff.

Replaces direct Baseten SDK integration for conversational/text-generation tasks. Requires huggingface_hub >= 1.26.1 (Python) or @huggingface/inference (JS) and HF_TOKEN. Ready now: drop-in replacement using OpenAI client with router.huggingface.co endpoint. Start with DeepSeek V4 Flash; support for other task types rolling out.

• “Baseten is now a supported Inference Provider on the Hugging Face Hub”
• “Baseten is launching support for conversational and text-generation tasks on Hugging Face”
• “huggingface_hub (>= 1.26.1) for Python and @huggingface/inference for JavaScript”
• “you'll only pay the standard provider API rates. There's no additional markup from us; we just pass through the provider costs directly”
• “PRO users get $2 worth of Inference credits every month”

Ling 3.0 Tiny launches free on Vercel AI Gateway

1.3B active parameters, 256K context, native function calling — swap model string to `inclusionai/ling-3.0-tiny-free` in AI SDK for responsive agents.

Replaces Ling 3.0 Flash in the free tier with a smaller MoE model tuned for multi-turn conversation and function calling. Cuts latency for agent loops while preserving 256K context and 32K output tokens.

Ready now. Drop-in replacement via model parameter change in AI SDK. Free through 8/14, then paid. Requires no code restructuring if already on AI Gateway. Worth testing for cost-sensitive agent workloads, but benchmark performance against Flash on your use case first.

• “Ling 3.0 Tiny is a MOE model with 7.9B total parameters and about 1.3B active per token”
• “a 256K token context window, and up to 32K output tokens”
• “The model is built for responsive agents, instruction following, and multi-turn conversation, with native function calling and prompt caching”
• “set `model` to `inclusionai/ling-3.0-tiny-free`”
• “free to use till 8:00am PT on 8/14”

Cloudflare AI Search adds multi-instance queries, public endpoints

Index distributed data sources (docs, blogs, APIs) and search across them with a single API call or zero-code public endpoint; embedding and reranking cost nothing with default models.

Replaces manual stitching of Workers AI + Vectorize + R2 with integrated ingestion and retrieval. Agents now query current, cited docs instead of relying on stale training data or slow full-page web fetches.

Ready now in beta (free until billing enabled). Requires wrangler CLI for setup or zero code if you use public endpoints. Worth trying if you're building agent-facing search or migrating from DIY vector pipelines; the namespace-level multi-instance binding is the key win.

• “embedding and reranking are free when you use the default models”
• “Point AI Search at a site, or set of sites, and it handles crawling, ingestion, embedding, and retrieval”
• “res.chunks come back cited and tagged with the instance they came from”
• “you immediately get /search and /mcp endpoints that query every instance, with no auth and nothing to deploy”

MCP protocol drops stateful sessions, simplifies server deployment

MCP 2026-07-28 removes session handshakes and open streams, letting servers run stateless on Workers instead of Durable Objects—each request now carries its own protocol context.

Eliminates sticky sessions, stream management, and deployment complexity for agents talking to external services. Developers can now deploy MCP servers as simple request-scoped functions instead of managing stateful infrastructure.

Replaces McpAgent + Durable Objects for most MCP deployments; requires SDK migration (TypeScript/Python/Go/C# updated). Breaking change for elicitation (now Multi Round-Trip Requests), but ready now—Cloudflare customers have validated against production traffic. Worth migrating if you control both client and server.

• “MCP is now a fully stateless protocol”
• “The latest MCP 2026-07-28 specification was released last week”
• “MCP servers can now run in just a Worker, no stateful infrastructure needed”
• “Each request carries the protocol version, client identity, and client capabilities it needs”
• “Customers and partners have used the release candidate on Cloudflare before the specification was finalized, giving us confidence that the migration path from McpAgent to the new createMcpHandler works with production traffic”

Data Point

Opus 5 reaches 24% on long-horizon coding benchmark

SlopCodeBench reveals multi-checkpoint progressive requirements expose model drift: Opus 5 accumulates defects steadily after initial wins, hitting regression on 93% of written code lines.

Standard benchmarks divulge entire problem upfront; SlopCodeBench forces iterative evolution with inherited state, exposing whether your LLM can maintain a codebase without human steering across multi-step requirements.

This replaces abstract code-quality metrics with concrete regression testing. Requires: running your own checkpoint-based evals if you care about real codebase maintenance. Ready now as a diagnostic tool, but Opus 5's 24% strict pass rate signals no model yet runs unattended on multi-step engineering work.

• “Opus 5 got a 24% on the small subset of the benchmark”
• “THERE ARE NO GOOD BENCHMARKS for a model's ability to maintain codebase quality”
• “each challenge in SlopCodeBench has multiple checkpoints - the model doesn't know the whole problem up front, it has to evolve the codebase over time as new requirements are divulged”
• “every dollar bought correctness. nobody bought enough of it.”
• “For all 9 test runs, none of the models made it to the end of any challenge with everything passing, even on the problem marked as easy difficulty”