Managed runtime architecture separates organization-wide decisions from the isolated environment where a specific agent run executes. The pattern is established in distributed systems even though the exact MIR terminology remains emergent. Current vendor platforms expose parts of this pattern. Source: AWS documentation Source: Google Cloud documentation
Key takeaways
- The control plane should not depend on the model to enforce identity or authority.
- A runtime instance needs an explicit owner, configuration, lifecycle, and termination path.
- Cross-plane evidence connects governance intent with actual execution.
Control-plane responsibilities
- Identity, tenancy, ownership, and delegation.
- Policy distribution and versioning.
- Runtime profiles, model routing, budgets, and region constraints.
- Provisioning, health, reconciliation, suspension, and termination.
- Organization-wide telemetry, incident response, and evidence retention.
Runtime-plane responsibilities
- Isolated workspace and process boundary.
- Scoped memory projection and task context.
- Model adapters and typed tools.
- Short-lived credentials and controlled outbound network access.
- Checkpoints, artifacts, local telemetry, and cleanup.
The runtime plane should receive only the authority required for the active objective. Long-lived administrative credentials and hidden cross-tenant memory are incompatible with this boundary.
Lifecycle state machine
- created
- provisioning
- running
- awaiting approval
- recovering
- completed
- failed
- terminated
State transitions should be validated, timestamped, attributable, and recoverable. “Awaiting approval” is a first-class state rather than an unstructured pause inside a chat transcript.
Operational boundaries
| Concern | Control plane | Runtime plane |
|---|---|---|
| Policy | Defines and versions rules | Evaluates local decisions and enforces outcomes |
| Secrets | Authorizes issuance | Receives short-lived scoped credentials |
| State | Tracks desired lifecycle | Maintains active task state and checkpoints |
| Telemetry | Aggregates and governs retention | Emits traces, events, metrics, and artifacts |
| Termination | Orders stop and retention policy | Revokes access, seals evidence, removes workspace |
Source record
References
- Amazon Bedrock AgentCore Runtime Developer Guide Primary source
Amazon Web Services. AWS. Published Current documentation; last reviewed 2026-06-20 UTC. Vendor documentation.
- Gemini Enterprise Agent Platform Overview Primary source
Google Cloud. Google Cloud. Published Current documentation; last reviewed 2026-06-20 UTC. Vendor documentation.
- AI Runtime Infrastructure Primary source
Christopher Cruz. arXiv. Published 2026-03; last reviewed 2026-06-20 UTC. Research paper.
OpenTelemetry project. Cloud Native Computing Foundation. Published Current specification repository; last reviewed 2026-06-24 UTC. Official specification.