fx joins Vercel’s AI harness layer

// 2h agoPRODUCT UPDATE

fx joins Vercel’s AI harness layer

ANNOUNCEMENT PRODUCT GITHUB PRODUCT HUNT X

Vercel’s lightweight, open-source coding agent fx now plugs into AI SDK’s HarnessAgent through the official @ai-sdk/harness-fx adapter, using ACP. Developers can embed fx through the same interface used by other supported coding harnesses.

// ANALYSIS

This is more important as plumbing than as a new agent feature: Vercel is making coding harnesses interchangeable, while fx brings a notably small native runtime to the stack.

–The ACP-based adapter removes the need for bespoke integration code.
–HarnessAgent standardizes generate and stream flows across coding agents.
–fx’s Zig implementation, small binary, model agnosticism, and MCP support make it attractive for embedded or sandboxed workloads.
–The project remains experimental, so teams should expect breaking changes and validate capability parity before production adoption.

// TAGS

fxai-codingcoding-agentagentopen-sourceclisdk

DISCOVERED

2h ago

2026-09-01

PUBLISHED

5h ago

2026-08-31

RELEVANCE

9/ 10

AUTHOR

vercel_dev

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

BENCHMARK43m ago

Celeris-1 Magnus Tops GPT-5.6 Sol in τ³-bench

Celeris-1 Magnus reportedly outscored GPT-5.6 Sol on τ³-bench while delivering faster response times than GPT-5.6 Luna. The result highlights Celeris’ low-latency positioning for tool-using AI agents.

OPEN SOURCE1h ago

Qwen3.8-27B EXL3 unlocks 200K local context

An experimental EXL3 quantization makes Qwen3.8-27B practical on consumer GPUs, including RTX 3090, 4090, and 5090 cards. Paired with DFlash2 and NVFP4 KV caching, the deployment kit targets roughly 200K–262K-token contexts within 24GB of VRAM.

INFRA2h ago

Qwen3.8-27B Runs at 417 Tok/s on 3090

A community vLLM recipe fits Qwen3.8-27B on one 24GB RTX 3090 using W4A16 quantization, requantized embeddings, and inference tuning. The open-weight multimodal model supports long-context coding and agent workflows, while the recipe reports 417 tokens per second in batched tests.