AWS AgentCore Runtime V2 brings snapshot cold starts and elastic memory

AWS shipped a second-generation runtime for Bedrock AgentCore, its managed hosting layer for agents. Runtime V2 restores agent sessions from snapshots instead of booting fresh, and manages memory elastically, releasing it when agents idle.

Cold starts are a key pain for agents that spin up per request or per user session. Developers praised the result: P75 cold starts of about 2 seconds compared with around 30 seconds previously, and lower bills because elastic memory reclaims unused capacity. AWS also introduced Web Answer, which returns cited, synthesized answers, complementing the InvokeWebSearch API that went GA a day later.

AgentCore is now a large platform: AWS says tens of thousands of customers use it, and Q3 Bedrock token volume rose nearly 17X year over year. The broader AgentCore push this week included guidance on defending persistent memory against poisoning through multi-layer write validation, a seven-layer reference architecture, and CrowdStrike integrating Falcon Guardian with AgentCore Guardrails. Competing hosted agent runtimes come from Google, Microsoft and newer entrants like Mistral Studio's managed agents.

Caveats: the 2s P75 figure comes from developer reports rather than an AWS benchmark, and tail latency (P99) matters for interactive agents. Watch for pricing changes tied to elastic memory.