Agent Platform: Inference Broker — hot-pluggable LangGraphs, not "logic layer"

The extensibility mechanism is a hot-pluggable LangGraph behind each
alias (same pattern used elsewhere in the ecosystem), not a fixed
"logic layer" of built-in behaviors. Rate-limit fallback, size-based
routing, format reshaping, content-aware dispatch are typical patterns
— but because each alias is a real graph, the range is open.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: rmdevpro

open-source/agent-platform.html

@@ -171,7 +171,7 @@ <h2>How they fit.</h2>
       <h2>Coming next.</h2>
       <p>Additional agentic services on the roadmap, designed to extend the platform. Concept-stage or early build, not yet shipping.</p>
       <ul>
-        <li><b>Inference Broker</b> — a single proxy in front of every model call across the ecosystem, local or cloud. Clients invoke a named alias; the alias either points straight at a specific model, or runs through a logic layer that shapes and redirects the call. That logic can fall back to a secondary model on a rate-limit error, send small requests to local inference and large ones to the cloud, reshape an OpenAI-format call for an Anthropic endpoint, or dispatch by content — code authoring to one model, code testing to another. Like the Context Broker, its Imperator can metacode new aliases, routes, and transformation logic.</li>
+        <li><b>Inference Broker</b> — a single proxy in front of every model call across the ecosystem, local or cloud. Clients invoke a named alias; the alias either points straight at a specific model, or runs through a hot-pluggable LangGraph that can do whatever the author wires it to do. Typical patterns include rate-limit fallback, size-based local-vs-cloud routing, format reshaping between providers (OpenAI → Anthropic), and content-aware dispatch (code authoring to one model, code testing to another) — but because each alias is a real graph, the range is open. Like the Context Broker, its Imperator can metacode new aliases and graph plugins.</li>
         <li><b>Conversation Broker</b> — agent-to-agent messaging on an email-style model: addressed communication, threading, participant history, delivery notification, loop protection. Internal, controlled, no SMTP required.</li>
         <li><b>Agentic File System</b> — intent-based filesystem management. Replaces POSIX permissions with a reconciliation between policy intent and action intent. Every mutation carries a stated reason; policies are expressed in terms of purpose, not just path patterns.</li>
         <li><b>Development Cache</b> — a shared cache for agentic development work.</li>