Google Cloud engineers describe changes to Envoy aimed at the demands of AI agents, including token-budget controls, Model Context Protocol support and routing designed to reuse model cache. The practical shift is towards a network proxy that can help manage the cost and traffic of agent systems, not just pass requests along.
Watch Desk analysis
What happened
In a Google Cloud engineering post, the team outlines native parsing for model wire formats, a tokenomics engine for budget enforcement, and KV-cache prefix affinity. It also describes MCP support and integration with Agent Substrate, including scale-to-zero traffic steering and egress perimeter defence. The Google Cloud post sets out the architecture; it does not provide comparative performance results.
Why it matters
Agents can make repeated model calls, move data between tools and consume tokens while a task unfolds. Budget controls and cache-aware routing address two practical parts of running that workload: keeping spending within bounds and avoiding needless repetition. MCP support also puts a widely used agent communication protocol into the networking layer.
That makes the proxy part of the AI infrastructure conversation. For operators, the appeal is a chance to manage model traffic and egress alongside ordinary network routing, rather than treating every agent call as an isolated application concern.
Our read
This is a useful blueprint for the plumbing behind AI agents, with several concrete mechanisms rather than a cloud-sized promise to “transform” everything. The open question is how these pieces perform in real deployments, and how portable the approach is beyond Google Cloud’s own stack. Networking has always been where ambitious software meets the bill.
What to watch
- Whether the described features are available to operators now, and under what conditions.
- How token-budget enforcement and cache affinity affect cost and latency in production.
- Whether MCP support and the Agent Substrate integration work across different model providers.
Discussion spark: Should model-budget controls and agent traffic policies live in the network layer, or should application teams keep that responsibility?
Sources and evidence
- Building an agentic stack in Envoy: routing models, tokens, and MCP (18 November 2026, 00:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.