Summary

The day's five signals span the full agent stack — the model, the coding agent, its plugin ecosystem, the compute it acts through, and the governance around it. Meta appeared twice: it open-sourced Muse Glimmer, a 30B Apache-2.0 multimodal model tuned for always-on local agents that runs offline on a single consumer GPU, and launched Muse Code, a beta terminal coding agent built for large repos that plans, writes, and validates code against Claude Code, OpenAI Codex, and Cursor. OpenAI pushed extensibility cross-vendor with Codex 0.147.0, adding portable Agent Plugins whose marketplaces span Amazon Bedrock and even rival Claude Code, plus a --approve-for-me flag and MCP 2026-07-28 support. Cloudflare reframed where agent browsing runs with Kitesurf, a from-scratch Rust/WebAssembly browser engine executing inside Workers V8 isolates. And Palantir widened governed model choice, routing DeepSeek V3.2, Z.ai GLM 5, and Moonshot Kimi 2.5 into AIP via AWS Bedrock. Together they sketch a maturing, increasingly open, edge-native, and governance-aware agent stack that every major platform is now contesting.

Key themes

  • Agents are moving local and open. Meta's Apache-2.0 Muse Glimmer packs a 30B multimodal agent into an 18-20 GB (4-bit) footprint that runs offline via llama.cpp and Ollama on a single consumer GPU, cutting inference cost and data exposure and challenging the assumption that useful agents need hosted frontier APIs.
  • The terminal coding-agent race is now contested by every major lab. Meta's Muse Code enters against Claude Code, OpenAI Codex, and Cursor, betting on repo-scale context over per-file assistance — the enterprise pain point where single-file assistants break down.
  • Coding-agent extensibility is going cross-vendor. OpenAI Codex 0.147.0 ships portable Agent Plugins whose marketplaces include Amazon Bedrock and even rival Claude Code, plus MCP 2026-07-28 support — favoring write-once plugins and lower lock-in over walled gardens.
  • Agent infrastructure is collapsing onto the edge. Cloudflare's Kitesurf runs a purpose-built Rust/WebAssembly browser inside Workers isolates, co-locating browsing with agent compute to make per-agent web navigation cheaper and more scalable than headless Chromium.
  • Enterprise platforms are competing on breadth of governed model choice. Palantir routed frontier Chinese open and commercial models (DeepSeek V3.2, Z.ai GLM 5, Moonshot Kimi 2.5) into AIP via AWS Bedrock, giving regulated customers long-context reasoning options without leaving AIP's permission and audit controls.

Notable items

  • Meta open-sources Muse Glimmer (Aug 10) — a 30B Apache-2.0 multimodal model distilled from Muse Spark, tuned for offline local agents (coding, function calling, file/schedule management, multi-step reasoning with failure recovery); 4-bit quantization fits it, KV cache, encoder, and drafter into a 24-32 GB VRAM budget, runnable on Hugging Face via llama.cpp and Ollama. (media; impact: medium)
  • Meta launches Muse Code (Aug 5) — a beta terminal coding agent designed for large repos that plans changes, writes code, and validates results, positioning Meta against Claude Code, OpenAI Codex, and Cursor's CLI. (media; impact: medium)
  • OpenAI Codex 0.147.0 (Aug 7) — introduces portable Agent Plugins (manifests, workspace publishing, cross-catalog search) with new marketplaces for Amazon Bedrock and Claude Code, a --approve-for-me CLI flag, and MCP 2026-07-28 protocol support. (repo/primary; impact: medium)
  • Cloudflare launches Kitesurf (Aug 7) — a Rust/WebAssembly browser engine (Blitz rendering, Firefox's Stylo CSS parser, Boa JS engine) running inside Workers V8 isolates; passes 215,000+ web platform tests and is free in beta via Browser Run, extending Cloudflare's push to collapse agent infrastructure onto one edge platform. (media; impact: high)
  • Palantir AIP expands its governed model catalog (Aug 4) — adds DeepSeek V3.2 (164K ctx), Z.ai GLM 5 (256K multimodal), and Moonshot Kimi 2.5 (200K ctx) via AWS Bedrock on non-georestricted US/EU/UK/JP enrollments with client-side tool use, under AIP's permission and audit controls. (primary; impact: low)

Source coverage

Source rows used: 5