# AgentCore Runtime Instances — adoption decision matrix

Use after reading the [Runtime Instances GA field guide](https://www.factualminds.com/blog/amazon-bedrock-agentcore-runtime-instances-ga-2026/). Score each row **0–2** (no / partial / yes). Sum ≥ 14 → pilot Runtime Instances on one agent this sprint. Sum 8–13 → keep microVM Runtime and re-score at next architecture review. Sum < 8 → stay on default serverless microVM Runtime.

| # | Question | 0 | 1 | 2 |
|---|----------|---|---|---|
| 1 | Agent sessions need wall-clock longer than **8 hours** (docs, research loops, multi-day batch agents)? | Sessions < 1h | Occasionally 2–8h | Routinely need up to **14 days** |
| 2 | Workload needs **GPU**, high-memory, or a specific EC2 family microVM cannot express? | CPU-only, fits microVM defaults | Memory-heavy CPU | GPU / specialized hardware required |
| 3 | Sustained compute (hours of active CPU/GPU per day) vs bursty chat turns? | Burst chat, seconds per turn | Mixed | Sustained resource-intensive loops |
| 4 | Primary Region is in the Aug 6, 2026 GA list (us-east-1/2, us-west-2, ap-south-1, ap-southeast-1/2, ap-northeast-1, eu-central-1, eu-west-1)? | Wrong region | DR region only | Primary in GA region |
| 5 | Team accepts **EC2 instance hours + AgentCore management** as two Cost Explorer lines? | Needs single serverless bill | Aware, not modeled | Modeled both lines |
| 6 | Can attach a **capacity provider** (instance types) without rewriting invoke/deploy path? | Needs custom host OS control | Partial | Same InvokeAgentRuntime / deploy APIs |
| 7 | Observability baseline exists (session duration, active vCPU/GB-seconds, tool latency, idle capacity)? | No metrics | Dashboards only | Alarms + session length histogram |
| 8 | FinOps owner will compare **$/successful session** (not just EC2 $/hour) after a 72h canary? | No owner | Ad-hoc | Tagged canary + CE report |
| 9 | Failure mode reviewed: always-on GPU for short chat agents → idle burn? | Not reviewed | Discussed | Explicit reject criteria written |
| 10 | Security/compliance OK with AgentCore-managed EC2 lifecycle (patching, scaling) vs DIY EC2 fleet? | Must own host patching | Hybrid | AgentCore lifecycle acceptable |

**Quick rule:** Default to **microVM Runtime** for sessions ≤ 8 hours that need fast startup. Choose **Runtime Instances** when session length, specialized hardware, or sustained load makes microVM the wrong unit of compute — not because “EC2 feels more production.”

**Related:** [AgentCore production guide](https://www.factualminds.com/blog/amazon-bedrock-agentcore-production/) · [12-component pricing](https://www.factualminds.com/blog/amazon-bedrock-agentcore-pricing-12-components/) · [pricing calculator](https://www.factualminds.com/tools/amazon-bedrock-agentcore-pricing-calculator/)