The Agent Stack Is Reshaping: Muse Glimmer, Claude Code Sessions, and GitHub Models Gone

Three agent-layer developments landed simultaneously on August 10, and together they signal a structural shift: the tooling layer is consolidating around open, interoperable, locally-runnable agents while hosted convenience layers are quietly disappearing.

Why it matters

Each story is notable alone. Together they point at the same thing: the ad-hoc, API-stitched agent stack that most teams built over the past two years is being replaced by something more deliberate.

• Meta's Muse Glimmer is an open-source multimodal agentic model designed to run locally. It handles vision, text, and tool use without a cloud round-trip. For teams with data-residency constraints or latency requirements, this is the first credible open alternative to hosted multimodal agents.
• GitHub Models is retired, as Simon Willison discovered when his CI pipeline broke. The retirement brownout is over; the service is gone. Any workflow that used GitHub Models for free or low-friction LLM access needs a replacement now.
• Claude Code now lets sessions communicate with each other on macOS, enabling agent-to-agent coordination within a single developer environment. This is a small surface area change with large architectural implications for multi-agent coding workflows.
• OpenAI's new Agent Plugins standard, released alongside GPT-5's one-year mark, formalizes how agents expose and consume capabilities across providers. It is a direct play for becoming the interoperability layer across the ecosystem.

The agent stack is no longer a set of API calls. It is becoming an infrastructure layer with its own protocols, local runtimes, and session models.

What changes in practice

• Teams using GitHub Models for prototyping or CI-based LLM evals need to migrate to a direct provider API or a self-hosted alternative immediately.
• Local-first agent architectures are now viable for multimodal tasks. Muse Glimmer removes a key blocker for on-device or air-gapped deployments.
• Multi-agent orchestration in Claude Code gains a native coordination primitive on macOS, reducing the need for custom message-passing scaffolding.
• OpenAI's Agent Plugins standard will pressure other providers to publish compatible interfaces. Teams designing agent tool schemas should watch the spec closely before locking in a proprietary format.

How to use it

1. Audit your GitHub Models dependencies today. Search your codebase and CI configs for github.com/marketplace/models or the GITHUB_TOKEN-based inference endpoint. Swap them for direct provider keys or a local model before the next pipeline run.
2. Evaluate Muse Glimmer for your multimodal agent tasks. Pull the model from HuggingFace, run it against your existing eval set, and benchmark latency versus your current hosted solution. Local wins on privacy; check whether it wins on quality for your use case.
3. If you run multi-agent Claude Code workflows on macOS, prototype with inter-session messaging now. Design your agent graph around native session coordination rather than building a custom broker you will have to throw away later.
4. Read the OpenAI Agent Plugins spec before finalizing any tool-schema design. If it gains adoption, building against it now costs little. Retrofitting later costs a lot. Check 9to5Mac's coverage for the initial spec summary.
5. Track the Agents and Open Source beats together. The interesting moves are happening at the intersection, not in either lane alone.

The teams that will be slowest are the ones still treating agent tooling as a convenience layer. It is becoming load-bearing infrastructure, and today's signals confirm the transition is already underway.