leejk/ jk lee
← Landscapes

Landscape · 2026.05

Agent Governance: The Chapter After Standardization

Registry, hosting, gateway, audit — the operational layer arriving on top of the standard, all at once

·
#agent-governance#agent-registry#mcp#governance#landscape

Immediately after MCP became standard — over a span of weeks — both major clouds and both major model providers entered the same space simultaneously. AWS Agent Registry, Microsoft Agent 365, Anthropic Claude Managed Agents, the OpenAI Agents SDK overhaul, and the central registry on the official MCP 2026 roadmap. All five point in the same direction: an operational/governance layer on top of the standard.

TL;DR. Right after MCP became standard, the operational layer that sits on top of it appeared all at once within the single month of April. AWS Agent Registry · Microsoft Agent 365 · Claude Managed Agents · the OpenAI Agents SDK major update · the official MCP roadmap — five entrants into the same space. Catalog, hosting, gateway, and audit are the four axes, and the breaking points differ by axis.

What Agent Governance Is

Definition — a shared-responsibility scheme for discovering, approving, executing, and auditing standardized agent resources (agents, MCP servers, skills). Not a single platform but a partition of operational responsibility.

Unpacked a bit, governance is a divided answer to four questions:

  • Where is it — Registry / Catalog
  • Where does it run — Managed Hosting
  • How is it reached — Gateway / Routing
  • What did it do — Audit / Observability

The 4-axis table below shows which products embed these four questions, and how. Even with five products showing up, there are only four questions to answer — which is why they're comparable under the same labels.

Why It Appeared Now

Three things crossed a threshold at once:

  1. The explosion in the number of resources. Public MCP servers crossed 10,000 in March 2026, and over April–May, one-person signature skill packs entered trending one after another. Add in-house homegrown agents and MCP servers and the cardinality of connectable resources has already passed what one person can track. That's the substance of "agent sprawl."
  2. The standard didn't unblock operations. MCP standardized how to talk but didn't resolve who bears operational responsibility. The three primitives over JSON-RPC are just a connection protocol — which server to trust, who has the right to kill it, where the call logs land are all problems outside the standard. The "the remaining problem is operations" that the MCP index piece concluded with is exactly this spot.
  3. Both major clouds and both major model companies entered in the same month. That AWS, Microsoft, Anthropic, and OpenAI all shipped answers in April 2026 is no coincidence. It's a signal that the operational vacuum right after standardization became visible to all four at once.

In one line: governance is the next chapter of the MCP standard. The standard solved connectability, so the operations on top of it must be solved by a separate layer — and that layer is appearing right now, all at once.

What Appeared at Once

Within the single month of April 2026, both major clouds and both major model companies put their answers to agent sprawl — the proliferation of agents and MCP servers inside an organization — in the same place.

  • 2026-04-09 — AWS Agent Registry (under Bedrock AgentCore, preview). Bundles four record types — MCP, A2A, Skills, Custom — into a single catalog, plus lifecycle, approval, and CloudTrail.
  • 2026-04 — Anthropic Claude Managed Agents (public beta). Declares MCP servers directly in the agent definition, runs in a hosted sandbox.
  • 2026-04-15 — OpenAI Agents SDK major update. Sandbox, subagent, long-horizon harness; MCP-compatible.
  • 2026-05-01 — Microsoft Agent 365 GA. Registry sync with AWS and Google, automatic inventory and lifecycle.
  • Official MCP 2026 roadmap — a central registry (npm-style), a Tasks primitive (async long-running), stateless HTTP transport.

All five point the same way. It's the first cloud answer to the "remaining problem is operations" that the MCP index piece concluded with.

The Four Axes of the Governance Layer

Decompose the five signals into four axes and you get this.

Axis What it solves Representative signals
Registry / Catalog Discovery, dedup, approval AWS Agent Registry, MS Agent 365, the official MCP registry roadmap
Managed Hosting Isolation, lifecycle, delegated operational responsibility Claude Managed Agents, AgentCore Runtime
Gateway / Routing Federation, alias, permission filtering AgentCore Gateway, agentic-community MCP Gateway, A2A interop
Audit / Observability Call logs, trust-boundary verification CloudTrail integration, tool annotations (MCP blog, 2026-03), agent obs tooling

The four axes are separable responsibilities, and each vendor bundles them differently. AWS separated Registry and Gateway, whereas Anthropic Managed Agents handles hosting and lifecycle as one bundle. The more the partitioning differs, the faster friction surfaces under concurrent operation.

Where It Will Break

Now that the operational layers have appeared at once, next quarter's operating question narrows to three.

  1. The single-source decision. When the same resource (MCP server, skill, agent) exists simultaneously in several governance layers, which one do you treat as canonical. Local vs. cloud Registry, your own Gateway vs. a vendor Gateway, a homegrown skill vs. a Registry-registered skill.
  2. Auth/permission model coherence. Registry uses IAM or OAuth, Managed Agents has its own sandbox permission model, Gateway has yet another permission policy. The omissions and duplications when three permission models meet on one call path.
  3. Lifecycle conflict. Cloud governance enforces an approval workflow (e.g., only APPROVED is exposed), but the local stack assumes immediate change. Apply two lifecycles to the same resource at once and both are neutralized.

These three are the recurring breaking points of operations over the next six months.

Per-Axis Detail

References

Sub-documents