Quick Answer
Choose AgentCore runtime instances when an agent has sustained, resource-intensive, specialized-hardware, or long-running compute needs that fit a specific EC2 instance family. Choose the existing serverless microVM-based runtime when sessions fit its current limits and fast startup plus serverless isolation better matches short or variable work.
AWS announced runtime instances as generally available on August 6, 2026. The launch says runtime instances can support sessions up to 14 days, while the default microVM-based runtime is designed for sessions up to 8 hours. Runtime instances do not mean unmanaged EC2: the user defines a capacity provider and instance types while AgentCore handles provisioning, patching, scaling, and lifecycle management.
Runtime Choice
| Decision | Runtime instances | Serverless microVM runtime |
|---|---|---|
| Workload shape | Sustained or resource-intensive agents | Shorter or bursty sessions needing fast startup |
| Session limit in launch record | Up to 14 days | Designed for sessions up to 8 hours |
| Compute fit | EC2 families including GPU, memory, and compute optimized | Managed default compute surface |
| Capacity | Capacity provider and selected EC2 types | Serverless runtime behavior |
| Cost | AgentCore compute-management charge plus EC2 and dependencies | Current AgentCore serverless pricing and dependencies |
| Operations | AgentCore manages infrastructure lifecycle, but the team still owns workload configuration and validation | Fewer instance choices, but session, startup, quota, and cost gates remain |
Nine-Region Launch Boundary
AWS listed runtime instances in US East (N. Virginia and Ohio), US West (Oregon), Asia Pacific (Mumbai, Singapore, Sydney, and Tokyo), and Europe (Frankfurt and Ireland). That launch list is not a permanent global guarantee. Confirm the exact account, Region, service quota, EC2 family, accelerator capacity, networking, and model or tool dependencies.
Migration Test
Use one replayable agent session and compare:
- initialization and time to first useful action;
- session duration, checkpointing, interruption, retry, and idempotency;
- CPU, memory, accelerator, disk, network, and instance-family fit;
- concurrency, scaling, queue, warm capacity, and failure isolation;
- IAM role, network path, secrets, tools, data stores, and egress;
- patching, image or dependency updates, observability, incident response, and rollback;
- AgentCore management, EC2, accelerator, storage, network, logs, model, and tool costs.
Do not migrate because 14 days is larger than 8 hours. Migrate when the real session needs the longer boundary or specialized compute and passes performance, security, availability, and cost acceptance.
Run the AgentCore runtime instance fit planner before creating capacity.
Frequently Asked Questions
Are AgentCore runtime instances just self-managed EC2?
No. AWS says you choose a capacity provider and EC2 types while AgentCore manages provisioning, patching, scaling, and lifecycle management.
How long can AgentCore runtime instance sessions run?
The August 2026 launch states up to 14 days, compared with a serverless microVM runtime designed for sessions up to 8 hours. Verify current documentation and quotas.
Do runtime instances support GPUs?
AWS says capacity providers can specify GPU-accelerated, memory-optimized, and compute-optimized EC2 families. Confirm the exact Region, family, quota, capacity, price, and workload compatibility.
Are runtime instances cheaper than serverless?
Not automatically. AWS states that runtime instances add a compute-management charge to EC2 costs. Compare the full accepted-session cost for the same workload.
Official Sources
Source check: August 15, 2026. Recheck Regions, session limits, instance support, quotas, charges, and documentation before deployment.