# microVM Runtime → Runtime Instances canary checklist

Run against **one non-production agent** first. Assumes the agent already deploys on AgentCore Runtime (microVM) via Console, CLI, SDK, or API.

Source announcement: [AgentCore runtime instances GA — Aug 6, 2026](https://aws.amazon.com/about-aws/whats-new/2026/08/aws-bedrock-agentcore-runtime-instances-generally-available/).

## Pre-flight (45 minutes)

- [ ] Confirm Region is GA for Runtime Instances (us-east-1, us-east-2, us-west-2, ap-south-1, ap-southeast-1, ap-southeast-2, ap-northeast-1, eu-central-1, eu-west-1).
- [ ] Export baseline for the control agent (microVM): median/p95 **session duration**, active **vCPU-hour / GB-hour**, tool failure rate, cold-start / first-token latency if you track it.
- [ ] Score the [decision matrix](./runtime-instances-decision-matrix.md) — proceed only if sum ≥ 14 **or** you have a hard requirement (GPU / >8h sessions).
- [ ] Model AgentCore Runtime lines on the [AgentCore pricing calculator](https://www.factualminds.com/tools/amazon-bedrock-agentcore-pricing-calculator/) **and** rough EC2 On-Demand for the candidate instance type (separate Cost Explorer mental model).
- [ ] Write rollback criteria: e.g. idle EC2 capacity > 40% over 24h, or $/successful session > 1.5× microVM control.

## Canary setup (same day)

- [ ] Create a **capacity provider** with the minimum instance types needed (start CPU-only unless GPU is mandatory).
- [ ] Attach **one** canary agent runtime to that capacity provider; leave production agents on microVM.
- [ ] Tag resources `MigrationWave=agentcore-runtime-instances-canary` for Cost Explorer.
- [ ] Run representative sessions for ≥ **72 hours** (include overnight idle windows — idle burn is the failure mode).
- [ ] Keep invoke path identical (`InvokeAgentRuntime` / existing SDK) — do not change product APIs for the canary.

## Compare (end of canary)

- [ ] Session success rate and tool error rate vs microVM control (equal traffic shape).
- [ ] Wall-clock session completion for long jobs (did you need the 14-day window?).
- [ ] **Cost:** AgentCore management + EC2 hours for canary vs microVM Runtime compute for control, normalized to **$/successful session**.
- [ ] Startup / schedule latency — Runtime Instances trade fast microVM startup for hardware fit; record the delta.

## Rollback triggers (stop canary if any fire)

- [ ] Idle EC2 capacity burns more than the microVM active-compute bill for the same successful session count.
- [ ] Session length never exceeds ~1 hour — you bought 14-day capability you do not use.
- [ ] GPU instance selected but GPU utilization < 15% averaged over canary window.
- [ ] Region or capacity-provider limits block scaling during a load spike.

## Promote or park

- [ ] If canary wins: document instance family + sizing; promote only agents that share that hardware profile.
- [ ] Keep a microVM path for bursty short-session agents — mixed fleets are supported.
- [ ] Schedule a 30-day FinOps review before buying Reserved Instances / Savings Plans on the EC2 side.
- [ ] Update architecture decision record with the scored matrix and CE screenshots.