> ../signals/2026-05-30.md
—— Signal one — 11% reach production ——
New enterprise data puts a number on what production teams have known for a year: 86–89% of enterprise agent pilots stall before reaching production scale. The primary reasons are not model quality. They are governance gaps, fragmented agent identity management, insufficient auditability, poor integration with existing systems, and vendor lock-in at the orchestration layer.
The 11–14% that do reach production share three characteristics that the stalled pilots lack. They defined their trust boundary before deployment, not after. They instrumented their agents with observable metrics from day one — not as an afterthought when something broke. They treated agent permission changes as production code changes, with review and justification required.
The practical implication: if your pilot is stuck, the bottleneck is almost certainly not the model. Audit your governance posture before you audit your prompts.
—— Signal two — The Five Eyes weigh in ——
The cybersecurity and intelligence agencies of the United States, Australia, Canada, New Zealand, and the United Kingdom published joint guidance titled "Careful Adoption of Agentic AI Services," addressing security risks in agentic AI systems deployed in critical infrastructure and defense environments.
Joint Five Eyes guidance is not a research paper. It is a policy signal. When the same security posture appears in guidance from five intelligence agencies simultaneously, it moves from "recommended practice" to "what your compliance team will ask about in 18 months."
The guidance focuses on four areas: agent identity and authentication, the minimal footprint principle (agents should acquire only the permissions and data access necessary for the current task), human oversight preservation, and supply chain risk in MCP server and tool integrations.
The minimal footprint principle is the one that matters most for production engineering. It is the least-privilege principle applied to agents — and it is not the default behavior of any major agent framework.
—— Signal three — A2A becomes a Linux Foundation standard ——
The Agent-to-Agent (A2A) protocol, developed initially by Google and subsequently adopted by major vendors, has been transferred to the Linux Foundation for open governance. Every major agent framework has now implemented A2A alongside MCP.
What this means for production teams: A2A is no longer a Google initiative. It is infrastructure. If your agent stack communicates across vendor boundaries — which most enterprise deployments do — A2A is the canonical interoperability layer. Build your tool contracts assuming A2A as the transport. This is not a future consideration. It is the current state.
The practical pairing: MCP for agent-to-tool communication, A2A for agent-to-agent communication. Both are now stable, open standards under foundation governance. The protocol layer of agent infrastructure has settled. The engineering work is now at the application layer.
—— What to do with this ——
◆ Signal I: Audit your pilot against three questions — do you have a documented trust boundary, observable metrics from day one, and a permission change review process? If any of the three answers is no, you have found your bottleneck.
◆ Signal II: Read the Five Eyes guidance. Not for compliance — for the minimal footprint checklist. It is the clearest public articulation of production-grade least-privilege for agents that has been published.
◆ Signal III: If you have multi-agent systems communicating across vendor boundaries, verify your A2A implementation version. The Linux Foundation transfer may trigger versioning changes in foundation-maintained SDKs over the next 90 days.
—— End of signal ——
◆ The pilot-to-production gap is a governance gap, not a model gap.
◆ Five Eyes guidance is a policy signal. Treat it as one.
A2A + MCP as stable standards means the protocol layer has settled. The work is now at the application layer.
— ORBIRESEARCH