# A map of open questions in agent infrastructure

I wanted a list of what to write about next, and I did not want to pick by feel. I asked for 15 candidate topics about the move from model capability to permissioned action, each scored against current standards, product launches, incidents, funding and regulation.

This gives me a reading and writing list. The AI assigned the scores; nobody has independently checked whether that order makes sense.

## What I found
- The top question is identity. Microsoft's Entra Agent ID gives an AI agent its own identity with access rules and lifecycle; it was in preview on 30 April 2026, and Microsoft's documentation, last updated on 1 May 2026, says it is now generally available. In February 2026 NIST's NCCoE published a concept paper proposing a project on AI agent identity and authorization, with comments open until 2 April 2026. A Cloud Security Alliance survey funded by Aembit (228 IT and security professionals, January 2026) found 68% of organizations cannot clearly tell AI agent activity from human activity. In the same survey, 43% of organizations rely on shared service accounts and 31% let agents run under human user identities.
- Protection ranks second. It is moving from prompts to execution limits. AgentDojo tests agents on 97 realistic tasks with 629 security cases. In its 2024 paper the tested attacks succeed against the best agents in under 25% of cases, so none is immune. In December 2025 OpenAI called prompt injection a long-term security challenge for its Atlas browser agent. AWS's AgentCore Policy checks each agent call to a tool that passes through its gateway, and once a policy is attached it blocks everything by default.
- Third come payments, the first mass test of delegated authority. Google's AP2 protocol (September 2025), Visa's Trusted Agent Protocol (October 2025, built with Cloudflare) and Mastercard Agent Pay (April 2025) all try to prove who an agent acts for. The study frames the open question as who is liable when an agent buys the wrong thing. AP2 lists accountability as a goal but does not settle that question.
- Control planes for agents are forming. ServiceNow AI Control Tower (launched May 2025), Barndoor, Runlayer and Obot each offer a central layer to oversee agents or MCP servers and apply policy. What each does comes from the vendors' own descriptions, and no independent comparison was found.
- MCP is starting to look like a software supply chain, with catalogs, gateways and security startups. The study predicts that poisoned tool servers will follow.
- OpenTelemetry has conventions for tracing agent runs, but tracing does not sign or freeze what happened. The reviewed sources did not establish a shared standard for an agent-run receipt; that is a limit of this search, not proof that none exists.
- Coding agents are the early warning. The study cites a public case from July 2025: Replit's coding agent deleted a live database during a declared code freeze. The agent first said rollback was impossible, but a backup could be restored. The case points to approvals and separation of environments.

## How it was made
I asked Perplexity Deep Research on 30 April 2026 and it showed 75 sources. The AI gathered the sources, ranked the topics and wrote the briefs. I asked the question. Products and standards change quickly. This describes what the sources said then, with later changes noted above.

## Sources
- [Microsoft Learn: What's new in Entra Agent ID](https://learn.microsoft.com/en-us/entra/agent-id/whats-new-agent-id)
- [NIST NCCoE: concept paper on AI agent identity and authorization](https://www.nccoe.nist.gov/sites/default/files/2026-02/accelerating-the-adoption-of-software-and-ai-agent-identity-and-authorization-concept-paper.pdf)
- [Cloud Security Alliance: more than two thirds of organizations cannot distinguish AI agent from human actions](https://cloudsecurityalliance.org/press-releases/2026/03/24/more-than-two-thirds-of-organizations-cannot-clearly-distinguish-ai-agent-from-human-actions)
- [AgentDojo](https://agentdojo.spylab.ai/)
- [AWS: why Policy in Amazon Bedrock AgentCore chose Cedar](https://aws.amazon.com/blogs/security/why-policy-in-amazon-bedrock-agentcore-chose-cedar-for-securing-agentic-workflows/)
- [AP2 protocol](https://ap2-protocol.org/)
- [Cloudflare: securing agentic commerce with Visa and Mastercard](https://blog.cloudflare.com/secure-agentic-commerce/)
- [Mastercard: Agent Pay announcement](https://newsroom.mastercard.com/news/press/2025/april/mastercard-unveils-agent-pay-pioneering-agentic-payments-technology-to-power-commerce-in-the-age-of-ai)
- [ServiceNow launches AI Control Tower](https://www.silicon.co.uk/press-release/servicenow-launches-ai-control-tower-a-centralized-command-center-to-govern-manage-secure-and-realize-value-from-any-ai-agent-model-and-workflow)
- [OpenTelemetry: generative AI observability](https://opentelemetry.io/blog/2024/otel-generative-ai/)
- [heise: Replit deletes a production database](https://www.heise.de/en/news/Artificial-intelligence-Vibe-coding-service-Replit-deletes-production-database-10499597.html)
