Reference stack
The AI agent stack, layer by layer
Sixteen layers and the specification that covers each one today. On 21 September sandboxing and agent identity had nothing; a month later both have a young specification.
- Standard
- 2 / 16
- Emerging
- 10 / 16
- Experimental
- 4 / 16
- No standard
- 0 / 16
By plane
Interface
Operator interface
Emerging
AG-UI · Agent Client Protocol · MCP Apps
Human-in-the-loop
Emerging
MCP elicitation 2026-07-28
Orchestration
Control plane / orchestration
Experimental
—
Agent-to-agent
Standard
A2A 1.0.1
Evaluation
Experimental
—
Governance & audit
Emerging
TRACE · W3C PROV-O
Capability
Tool calling
Standard
MCP 2026-07-28
Model access
Emerging
Open Responses · OpenAI Responses API
Knowledge / retrieval
Experimental
—
Context & memory
Experimental
AGENTS.md · Agent Skills
Trust
Observability
Emerging
OTel GenAI semconv
Agent security controls
Emerging
OWASP ACS 0.1.0
Secrets brokering
Emerging
RFC 8693 · SPIFFE
Policy / authorisation
Emerging
MCP Authorization 2026-07-28
Agent identity
Emerging
draft-ietf-wimse-aims-00 · SPIFFE
Execution
Sandboxing / execution
Emerging
OpenSandbox · OCI Runtime Spec
Specifications and implementations
Method
Standard: a published specification under neutral governance, with independent implementations. Emerging: a specification exists, but it is young, partial or comes from one sponsor. Experimental: frameworks and products, no shared specification. Specifications checked against their primary source on 26 October 2026; implementations against their official pages on 23 October 2026.