MCP 2026-07-28 replaces a bidirectional stateful core with request/response, formalizes extensions, and hardens authorization; Claude support is rolling out.
Why it matters: For developers, the changes are intended to simplify deployment and scaling, add capabilities without changing the core protocol, and connect to enterprise identity systems without workarounds. The rollout caveat matters: support is not stated to be available across all Claude products yet.
Prompt-only workflow instructions leave an LLM agent in control of both procedure choice and step execution. COVENANT instead compiles the workflow and checks proposals before controller state changes or the graph advances.
Why it matters: For multi-step policies, the practical shift is from asking a model to remember the procedure to checking each proposal before the workflow advances. The reported results are promising for the evaluated cases, not a guarantee of broader performance.
OrchBench uses deterministic simulation to compare multi-agent orchestration plans without invoking worker agents, separating plan quality from other execution effects.
Why it matters: The simulator offers a lower-cost way to compare and diagnose orchestration plans. The reported findings also direct attention to preserving task-critical information, not simply increasing the number of agents.
MCP-based Agent Skills let .NET agents discover centrally hosted skills and load content on demand, avoiding per-agent copies while providing explicit bounds for archive extraction.
Why it matters: Teams can maintain shared policies and playbooks in one server rather than packaging copies into each agent. For archives, configure the stated limits to bound remote content before extraction; note that the API is experimental.