Short answer
AWS launched the next generation of Amazon Bedrock AgentCore Runtime on September 18, 2026, adding elastic memory reclamation and snapshot-based starts for serverless AI agents. AWS reports P75 cold starts of roughly 1.9–2.0 seconds in a narrow echo-agent test across 200 MB to 2 GB images, compared with 5.4–30 seconds on V1. That benchmark isolates platform startup; it does not measure model latency, tool calls, retrieval, application correctness, or end-to-end user experience.
Free workflow pilot checklist
Test the workflow before you buy the tool.
Get the buyer checklist, including task, owner, approval, fallback, and time-saved fields—plus one useful briefing a week.
What changed
V2 starts sessions with a smaller memory profile, pages in more memory as needed, and reclaims cold or released memory instead of billing against the session's high-water mark. It initializes a container once, captures a trimmed snapshot after health checks, and restores new instances from that snapshot. Teams opt in with the V2 platform version rather than receiving an invisible application migration.
Why it matters
Long-running and event-driven agents spend substantial time waiting on models, tools, people, or external systems. Reclaiming unused memory can better align infrastructure charges with active work, while more consistent starts can improve interactive responsiveness. Neither change fixes inefficient prompts, expensive inference, runaway tools, retries, or poor task design.
What to test before migration
Replay representative short, long, bursty, and idle sessions on V1 and V2. Measure p50/p95 startup and end-to-end latency, resident memory, GB-hours, model and tool cost, concurrency, snapshot safety, initialization side effects, failures, recovery, and regional availability. Keep a rollback path and compare total accepted-work cost—not the runtime rate alone.
Bottom line
AgentCore V2 is a meaningful infrastructure update for AWS-based agents, especially variable and long-lived workloads. Treat AWS's benchmark as a testable platform claim, not a promise that every agent becomes faster or cheaper.
Transparency
How this guide was checked
Editorial accountability
Who checked this guide
- Evaluation type
- Research-based verification
- Last materially checked
- Evidence
- 4 listed sources
Hands-on testing is identified explicitly. Research-based coverage uses cited product documentation and other named sources; it does not imply every paid plan was used. Read the full methodology.
Editorial basis
What this guidance is based on
- Editorial basis
- Source-led analysis
- Primary references
- 4
- Products covered
- 2
- Last checked
- 2026-09-22
Important limits
- • The performance figures are AWS measurements from a narrow echo-agent test.
- • Regional availability, compatibility, workload behavior, and total cost vary.
Sources and references
References cited by this article. Check the original source for scope, dates and current terms.
Frequently asked questions
What is AgentCore Runtime V2?
It is AWS's next-generation managed microVM runtime for AI agents, with on-demand memory, memory reclamation, and snapshot-based starts.
How fast are AgentCore V2 cold starts?
AWS reports a P75 of about 1.9–2.0 seconds in its specified echo-agent benchmark; real applications add model, network, retrieval, and tool latency.
Is AgentCore V2 automatically cheaper?
No. Reclaimed memory may reduce runtime GB-hours, but workload shape, the higher rate, inference, tools, storage, retries, and data transfer determine total cost.
How do developers enable V2?
AWS says to set the runtime platform version to V2 when creating or updating a runtime, then validate compatibility and rollback.
Tools mentioned in this article
Vapi
Developer infrastructure for composing and operating real-time voice agents
Vapi lets teams combine speech recognition, models, voices, telephony, tools, and observability, but layered per-minute cost, retention, consent, reliability, and escalation must be proven with real calls.
Read next
