# AgentCore Runtime microVM V1 → V2 adoption decision matrix

Use after reading the [Runtime V2 GA field guide](https://www.factualminds.com/blog/amazon-bedrock-agentcore-runtime-v2-ga-2026/). Score each row **0–2** (no / partial / yes). Sum ≥ 14 → canary `platformVersion=V2` on one microVM runtime this sprint. Sum 8–13 → stay on V1 and re-score when session length or image size changes. Sum < 8 → keep V1.

This matrix is for **microVM platform version V2**, not [Runtime Instances](https://www.factualminds.com/blog/amazon-bedrock-agentcore-runtime-instances-ga-2026/) (EC2 capacity providers).

Source announcement: [The new AgentCore Runtime is now available — Sep 18, 2026](https://aws.amazon.com/about-aws/whats-new/2026/09/new-agentcore-runtime-generally-available/). Platform details: [Runtime how-it-works](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-how-it-works.html). Snapshot rules: [Optimize for Runtime V2](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-v2-optimize.html).

| # | Question | 0 | 1 | 2 |
|---|----------|---|---|---|
| 1 | Cold starts hurt because the **container image is large** (hundreds of MB to multi-GB) or concurrency spikes create a start-time tail? | Image < 200 MB, starts already < 3s | Occasional 5–15s tails | P75 start tracks image size; AWS V1 range was **5.4–30s** |
| 2 | Sessions stay warm long enough for **unused memory to sit idle ≥ 120 seconds** after a spike (human wait, LLM I/O, tool round-trips)? | Median turn **3–5s**; reclaim never fires | Mixed 1–10 min | Multi-minute to multi-hour sessions with a memory spike then idle |
| 3 | You have modeled **V2 unit rates** ($0.1276/vCPU-hour, $0.0169/GB-hour consumption) against V1 ($0.0895 / $0.00945) on the [cost worksheet](./v1-peak-vs-v2-elastic-cost-worksheet.csv)? | Not modeled | Back-of-envelope | Worksheet filled; $/session compared |
| 4 | Primary Region is in the Sep 18, 2026 V2 list (`us-east-1`, `us-east-2`, `us-west-2`, `eu-west-1`, `ap-northeast-1`)? | Wrong region | DR region only | Primary in V2 region |
| 5 | Agent code is **snapshot-safe**: per-request IDs, clocks, and credentials live in the handler; Gateway tool catalogs are not frozen at startup? | Startup caches tokens / catalogs / `random` seed | Partially reviewed | Follows [optimize-V2](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-v2-optimize.html); BYO crypto uses snapsafe OpenSSL |
| 6 | `/ping` reports healthy only after init, and init finishes within **120 seconds** (V2 snapshot is taken on first healthy ping)? | Init often > 120s | Untested | Measured < 120s including model-weight load |
| 7 | Environment variables fit V2 caps (**1.5 KB** direct code, **2.5 KB** container vs **4 KB** on V1)? | Over 2.5 KB / 4 KB today | Close to cap | Well under cap |
| 8 | Team accepts **Console / CLI / SDK** to set `platformVersion` because CloudFormation and CDK cannot set it yet? | Must be CFN/CDK-only | Temporary CLI exception | CLI/SDK path approved |
| 9 | Observability baseline exists (session duration histogram, vCPU-hour, GB-hour, P75 start, tool errors)? | No metrics | Dashboards only | Alarms + Cost Explorer tags |
| 10 | Failure mode reviewed: short chat agents pay the **rate hike with zero reclaim**? | Not reviewed | Discussed | Explicit reject: median session < 120s and no cold-start pain → stay V1 |

**Quick rule:** Default **short-session chat** (turns that never reach 120s of unused memory) to **V1**. Choose **V2** when snapshot cold starts (AWS P75 **1.9–2.0s** vs V1 **5.4–30s**) or long idle-after-spike memory dominate the bill — not because the What's New page said "lower costs."

**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 (still V1 peak-memory rates, 2026-07-04)](https://www.factualminds.com/tools/amazon-bedrock-agentcore-pricing-calculator/)