Gate 1
Session shape
Record startup target, normal and maximum duration, idle windows, checkpoints, interruption tolerance, retries, and whether work can finish inside the current serverless session boundary.
Evidence-first AgentCore compute choice
Describe one real agent session, compute profile, Region, dependencies, failure behavior, security boundary, and budget. Generate a runtime-instance versus serverless acceptance plan without turning a longer session limit into a migration mandate.
Gate 1
Record startup target, normal and maximum duration, idle windows, checkpoints, interruption tolerance, retries, and whether work can finish inside the current serverless session boundary.
Gate 2
Measure CPU, memory, GPU or accelerator, disk, network, dependency, concurrency, and instance-family requirements with one replayable agent session.
Gate 3
Confirm account, Region, EC2 family and quota, capacity provider, IAM, networking, secrets, patching, scaling, observability, incident, and rollback paths.
Gate 4
Compare AgentCore management, EC2, accelerator, storage, transfer, logs, model, tool, idle, retry, and engineering costs per accepted session.
A sustained, specialized, resource-intensive, or longer-than-eight-hour session passes EC2 fit, Region, security, availability, and cost tests.
The workload fits the current microVM session boundary and fast startup plus variable demand outperform dedicated capacity for accepted sessions.
Session length or hardware points toward instances, but quota, capacity, networking, checkpoint, cost, or rollback evidence remains incomplete.
Replay the same session on both candidates; preserve inputs and versions; test interruption and resume; compare startup, p95 action latency, failures, compute utilization, scale behavior, security evidence, and full accepted-session cost; then exercise the rollback.
Official facts checked August 15, 2026 against the AWS GA announcement and documentation. Recheck live Regions, quotas, instance support, and pricing.