NEWS · COMPANIES · #3504
Amazon updates Bedrock AgentCore runtime for faster, larger, more efficient isolated agent sessions
Amazon announced next-generation updates to Bedrock AgentCore runtime, keeping per-session hardware isolation via Firecracker microVMs while improving startup, scale, and efficiency. Sessions now start up to 15× faster, boot in under 2 seconds independent of image size, scale up to 4× more compute (up to 8 vCPUs and 32 GB) with support for container images up to 10 GB, and run up to 68% more efficiently; customers report much shorter cold starts and reduced per-call costs in early tests.
KEY POINTS
- Amazon announced next-generation updates to Bedrock AgentCore runtime, keeping per-session hardware isolation via Firecracker microVMs while improving startup, scale, and efficiency.
- Sessions now start up to 15× faster, boot in under 2 seconds independent of image size, scale up to 4× more compute (up to 8 vCPUs and 32 GB) with support for container images up to 10 GB, and run up to 68% more efficiently; customers report much shorter cold starts and reduced per-call costs in early tests.
- Secure, low-latency, and cost-efficient isolated runtimes are a practical bottleneck for production AI agents; these updates remove performance and cost trade-offs while preserving hardware isolation.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
Amazon Bedrock AgentCore runtime (platformVersion “V2”) changes session startup from per-session container initialization to a snapshot-and-restore model: AgentCore prepares the container once, waits for a first healthy /ping, freezes the microVM, stores a snapshot, and restores that snapshot for each new session. Sessions continue to run in isolated Firecracker microVMs. V2 also reclaims memory sessions don’t use, supports build hooks, loads image file blocks lazily, and makes create/update operations take a few minutes while the snapshot is built.
WHY NOW
Excerpts state that agents are running longer and with more autonomy, increasing demands on secure, reliable, and cost‑efficient execution. As organizations move from a few agents to many concurrent sessions, balancing isolation, startup latency, and usage‑based cost becomes a practical bottleneck. V2 aims to make strong isolation compatible with faster, more elastic, usage‑based execution.
WHO IS AFFECTED
Directly affected: developers and customers using Amazon Bedrock AgentCore to deploy and run agents (AgentCore runtime is described as the managed compute layer). Indirectly affected: organizations scaling multi‑step, long‑running or interactive AI agents that require secure isolated runtimes.
CONFIRMED
- AgentCore runtime runs each agent session in its own isolated microVM. - V2 changes startup: AgentCore prepares the container once, waits for a first healthy GET /ping, clears OS file cache, freezes the microVM, stores its memory and disk as a snapshot, and restores that snapshot for new sessions. - Create or update operations now take a few minutes while the snapshot is built; the first healthy /ping must arrive within 120 seconds during snapshot creation. - V2 restores the snapshot into a new microVM for each new session; files from the image are loaded lazily (each block fetched on first read). - V2 reclaims memory sessions don’t use so billing tracks what the agent uses; the platform scales to zero when there’s no work and is billed for resources consumed. - AgentCore runtime is presented as a fully managed, serverless environment (managed compute layer) to deploy and run agents without building or maintaining infrastructure.
UNCERTAIN
Missing or not explicit in supplied excerpts: numeric performance and capacity claims in the event summary (e.g., “up to 15× faster”, “boot in under 2 seconds independent of image size”, “scale up to 4× more compute / up to 8 vCPUs and 32 GB”, “support for container images up to 10 GB”, “run up to 68% more efficiently”, and reported per‑call cost reductions). The excerpts describe the snapshot/restore approach, faster and more consistent startup, and memory reclamation, but do not provide those specific numeric measurements, aggregate customer metrics, or per‑call cost figures. Also not explicit: which specific agent architectures require code changes to get full V2 benefit (the excerpts say “some agents need restructuring” without enumerating them), and how V2 affects interactive experiences where even small latencies are critical beyond the qualitative statement that some experiences may still require prewarming.
WHAT TO WATCH
Observable signals to watch (from supplied excerpts): - Duration of create/update operations (snapshot build time reported as “a few minutes”). - Whether first /ping during snapshot creation meets the 120‑second requirement in practice. - Distribution of session cold‑start latency and its consistency after switching to V2 (measure session start time as described in the blog). - Per‑session memory usage and billing behavior showing reclaimed memory and usage‑based charges. - Frequency of lazy file block fetches and whether image size continues to dominate cold starts (V1 pulled the image per session; V2 loads blocks lazily). - Documentation and guidance updates or examples showing which agent patterns require restructuring or prewarming (the blog outlines patterns and build hooks).
WHY IT MATTERS
Secure, low-latency, and cost-efficient isolated runtimes are a practical bottleneck for production AI agents; these updates remove performance and cost trade-offs while preserving hardware isolation.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
AgentCore runtime runs each agent session in its own isolated microVM.
SUPPORTEDChecked 2026-11-13 · 1 supporting
PlatformVersion 'V2' prepares an environment once, takes a snapshot, and restores that snapshot for new sessions, which makes session start times faster and more consistent.
SUPPORTEDChecked 2026-11-13 · 1 supporting
V2 reclaims memory that sessions don’t use so billing tracks what an agent actually uses.
SUPPORTEDChecked 2026-11-13 · 1 supporting
Under V1, cold-start latency was dominated by pulling the container image, so cold start grew with image size.
SUPPORTEDChecked 2026-11-13 · 1 supporting
SOURCES & TIMELINE
1AI agents have moved past one-off prompts. They now run multi-step workloads for hours at a time. The length of task that frontier agents can complete on their own doubles about every four months ( METR ), and 40 percent of large enterprises are scaling AI agents, up from 27 percent a year ago ( McKinsey ). As agents run longer and with more autonomy, the focus shifts from what agents can do to how securely, reliably…
Agents today review code, process claims, and run unattended for hours. Many also sit behind interactive experiences where a person waits for the first response. Amazon Bedrock AgentCore runtime gives these agents a fully managed, serverless environment: each session runs in its own isolated microVM, the platform scales to zero when there’s no work, and you only pay for the resources your sessions consume. AgentCore…