How much faster is Gemini 3.7 Flash at code generation compared to 3.6?

Gemini 3.7 Flash achieves 43.6% on FrontierCode benchmarks versus 3.6 Flash's 34.4%. It also improves document reasoning from 22.0% to 34.0% on GDP.pdf, while costing half as much per token.

How do I switch my eve agent to GLM 5.2?

Set `model: "zai/glm-5.2"` in `agent/agent.ts` or run `eve set --model zai/glm-5.2`. Free access is available through August 27 via Vercel's AI Gateway; a one-line config change enables benchmarking.

What does the @ai-sdk/harness-acp package do?

@ai-sdk/harness-acp wraps the Agent Client Protocol, letting developers build harness adapters once against the protocol instead of writing runtime-specific code for each ACP-compatible package. Direct adapters still exist for Claude Code and Codex.

Can I use the Vercel v0 API in CI/CD pipelines?

Yes. The v0 API exposes the code-generation agent as infrastructure via REST with streaming agent actions, enabling integration into scripts, CI pipelines, and multi-agent systems without leaving the build environment.

What LLM version is required for llm-gemini Gemini 3.7 Flash support?

LLM 0.32 or higher is required. The plugin adds Gemini 3.7 Flash support with reasoning traces and server-side tool execution, reducing boilerplate for multi-step reasoning workflows.

Gemini 3.7 Flash: Coding Speed Breakthrough

Tool of the Week

Gemini 3.7 Flash launches with coding performance gains

3.7 Flash ships at half the 3.6 cost with measurable gains on code generation (43.6% vs 34.4% FrontierCode) and document reasoning (34.0% vs 22.0% GDP.pdf), available now via Gemini API.

Lower cost-per-token on production-ready code generation reduces inference spend for agentic workflows. Better first-pass accuracy means fewer retries in multi-step planning tasks, directly cutting developer iteration cycles.

Replaces 3.6 Flash for code/agent workloads. Requires no setup changes—drop-in swap via Gemini API. Ready now; introductory pricing valid through year-end. Worth migrating if you're already on Flash for coding or document processing.

“our most intelligent workhorse model yet for coding and agents”
“43.6% vs 34.4%”
“65.3% vs 49.0%”
“$0.75/1M input tokens and $3.75/1M output tokens”
“30.4% vs 17.0%”
“It thinks more diligently, putting in more effort into multi-step planning and tool calls”

gemini-apicode-generationcost-reductionagentic-workflowsbenchmarks

GLM 5.2 free for eve agents through August 27

Z.ai's 1M-token open-weights coding model is default on eve agents with free access until end of August via Vercel's AI Gateway.

Developers building eve agents get a capable coding model with extended context at zero cost during the trial window. Switching is a one-line config change, making it trivial to benchmark against your current setup.

Replaces your current eve agent model. Requires: one-line change to `agent/agent.ts` (set `model: "zai/glm-5.2"`) or `eve set --model zai/glm-5.2`. Worth trying now if you're already on eve; test before August 27 to evaluate quality before paid tier kicks in.

“GLM 5.2, the open-weights coding model from Z.ai with a 1M-token context window, is free for eve agents through August 27”
“New eve agents come with GLM 5.2 as their default model”
“After August 27, GLM 5.2 stays available on AI Gateway at standard provider rates”

llm-modelseve-agentscode-generationvercel-ai-gatewayfree-tier

AI SDK harness layer supports ACP protocol

@ai-sdk/harness-acp wraps Agent Client Protocol instead of specific runtimes, letting you adapt any ACP-compatible harness with minimal boilerplate.

Reduces harness adapter maintenance burden: instead of writing runtime-specific adapters for Claude Code, Codex, Pi, et al., you build once against the protocol. Direct adapters still exist for tight integrations, but ACP saves work when a harness only ships ACP-compatible packages.

Replaces need to write custom harness adapters for ACP-compatible packages. Requires npm package with ACP support, basic config (permission mapping, auth, env forwarding). Ready now—use it when direct adapters don't exist; prefer direct adapters for Claude Code and Codex.

“the AI SDK harness layer now supports any Agent Client Protocol (ACP)-compatible harness with HarnessAgent through the new @ai-sdk/harness-acp package”
“it lets you build an adapter for any harness that ships an ACP-compatible package”
“For Claude Code and Codex, prefer the direct @ai-sdk/harness-claude-code and @ai-sdk/harness-codex adapters over an ACP-based implementation”

ai-sdkharness-layerprotocol-abstractionacpadapters

Vercel v0 API now available for programmatic app generation

v0 API exposes Vercel's code-generation agent as infrastructure—invoke it from scripts, CI pipelines, or other agents via REST with streaming agent actions.

Moves AI-assisted development from interactive UI to composable infrastructure. Developers can chain v0 into deployment pipelines, multi-agent systems, or custom tools without leaving their build environment.

Replaces manual agent prompt-and-screenshot workflows with programmatic state management via chat IDs. Requires SDK integration and API key setup. Worth trying now if you're already building agent tooling or CI automation; overkill for one-off UI generation.

“Developers can send prompts through the API, have v0 generate and modify application files, run the resulting app in a Vercel Sandbox, and receive a preview URL that can be embedded into their own interfaces.”
“Developers can send follow-up messages using the same chat ID, allowing v0 to continue editing the existing application.”
“The API supports synchronous, asynchronous, and streaming requests, with streamed responses exposing individual agent actions such as file reads, edits, searches, Bash commands, and tool calls.”
“expose the application-building agent as infrastructure that developers can invoke from their own products, scripts, CI pipelines, or other agents”
“An agent can send a prompt and design system to the v0 API, receive a live preview in a Vercel Sandbox, iterate on the application, and deploy it when ready”

vercelai-agentsapicode-generationci-cd

Grok Build harness joins unified AI SDK interface

Grok Build is now available as `@ai-sdk/harness-grok-build`, letting you swap it for other coding agents through the standard `HarnessAgent` interface without changing application code.

Reduces lock-in to a single coding-agent runtime. If Grok Build underperforms or pricing changes, you swap the harness parameter instead of refactoring agent code.

Replaces direct Grok Build SDK imports with the standardized harness adapter. Requires `@ai-sdk/harness-grok-build` and minimal setup (three lines shown). Ready to use now if you're already on AI SDK v7.

“The AI SDK harness layer lets you run established coding-agent runtimes through one unified interface, so you can switch runtimes without changing your application code”
“@ai-sdk/harness-grok-build is the official harness adapter for Grok Build, built on top of the ACP harness adapter”
“The full supported list of harnesses is now: Claude Code, Codex, Deep Agents, Grok Build, OpenCode, Pi, with more coming soon”

ai-sdkgrok-buildharness-abstractionagent-runtimevercel

llm-gemini plugin adds Gemini 3.7 Flash support

llm-gemini now supports Gemini 3.7 Flash with reasoning traces and server-side tool execution via LLM 0.32 compatibility.

Developers using the LLM CLI can now access latest Gemini models with extended thinking capabilities and native code execution without client-side tool management. Server-side tools reduce boilerplate for multi-step reasoning workflows.

Drop-in upgrade for existing llm-gemini users. Requires LLM 0.32+. Worth adopting now if you use Gemini models via CLI; replaces manual API calls for reasoning + tool execution. One caveat: SVG rendering differs across browsers in generated images, test output in your target browser.

“This version of the plugin adds support for today's Gemini 3.7 Flash release, plus gemini-3.6-flash, gemini-3.5-flash-lite and two embedding models gemini-embedding-2 and gemini-embedding-001”
“you can now see reasoning traces and you can also enable server-side tools using this pattern”
“minimal, which was an option in 3.6 Flash, has been removed in 3.7”

geminillm-cliplugin-releasereasoning-modelstool-execution

Data Point

Multi-constraint LLM performance collapses past five

When stacking 5+ simultaneous constraints (schema, safety, reasoning), even top models drop below 50% pass rate due to multiplicative independent failures, not pairwise interference.

Production deployments routinely combine format, safety, and logic constraints. This quantifies exactly when your prompt engineering hits a wall and reveals that constraint interaction is multiplicative, not additive—changing mitigation strategy.

Doesn't replace anything yet; this is diagnostic data. Immediately relevant for teams shipping multi-constraint systems: stress-test your prompts/guardrails beyond 5 concurrent rules. The 2x structural-constraint penalty suggests routing complex reasoning and output format through separate calls or model passes may recover degraded performance.

“per-constraint pass rate decays gradually and predictably, while the chance of satisfying all k constraints collapses”
“a model passing individual constraints at ~41% at k=8 succeeds on all eight just 5.7% of the time”
“structural constraints lose 2x more baseline capability per added constraint than lexical ones”
“Reliable instruction following breaks down beyond 5-6 simultaneous constraints: probe-level success falls below 50% at 7 constraints for the strongest model”

constraint-compositionllm-reliabilityprompt-engineeringinstruction-followingbenchmark