AWS Cuts Bedrock AgentCore Cold Starts 15x With Firecracker Snapshots and Ships Native Web Search API
Amazon Web Services made its V2 AgentCore runtime the default on 12 November, cutting 2 GB container cold starts from 29.7 seconds to under 2 seconds alongside a native IAM-secured Web Search API.

By Aarav Mehta, Cloud & Enterprise Tech
, updated 6 min read
Amazon Web Services has rolled out a major architectural upgrade to Amazon Bedrock AgentCore, slashing session cold-start latency roughly fifteenfold via Firecracker microVM snapshots while introducing a standalone Web Search API built for enterprise agent pipelines.
Announced across back-to-back AWS disclosures between 10 November and 12 November 2026, the updates target two of the largest practical bottlenecks facing production AI agents: slow sandbox spin-up times for large dependency containers and brittle third-party search scraping wrappers.
How Snapshot Restoration Drops Cold Starts Below Two Seconds
Starting 12 November, AWS set the next-generation AgentCore runtime (platformVersion V2) as the default execution environment for all newly provisioned agent runtimes. Like its predecessor, V2 isolates every individual agent session inside a dedicated hardware-virtualised Firecracker microVM, preventing cross-tenant memory leaks when agents execute untrusted code or browse external files.
Where V2 changes the operational profile is initialisation. Instead of cold-booting a full container image and loading Python or Node.js frameworks from scratch on every new user session, AgentCore V2 initialises the runtime once and captures a warm memory snapshot. Subsequent sessions restore directly from that snapshot, cutting P75 cold-start latency for 2 GB container images from 29.7 seconds to under 2 seconds. For interactive developer coding agents, customer-service workflows, and financial triage pipelines, cutting nearly half a minute of boot overhead changes agent sessions from asynchronous background tasks into responsive web interactions.
AWS also expanded per-session resource envelopes and storage integration. Individual sessions can scale up to 8 vCPUs and 32 GB of RAM for heavy local data processing inside the microVM, while container image ceilings rise to 10 GB to accommodate specialised ML libraries, headless browsers, and local SDKs. Idle agents scale down to zero compute billing while dynamically reclaiming unused memory, paired with direct Amazon S3 mounts housed inside the microVM's trust boundary for multi-hour workflows.
Bringing Live Web Search Inside the AWS Trust Boundary
A day earlier, on 10 and 11 November, AWS made the AgentCore Web Search API (InvokeWebSearch) generally available. Previously, teams using Bedrock agents had to wire web queries through AgentCore Gateway and Model Context Protocol servers, or purchase external search API keys outside their cloud compliance perimeter.
With InvokeWebSearch, custom agent loops, data ingestion jobs, and AWS Lambda functions can query Amazon's operated web index through a single signed API call using standard AWS SDKs or the AWS CLI. Requests inherit standard IAM role permissions, CloudTrail audit logging, and VPC isolation via AWS PrivateLink, eliminating the need to manage separate third-party API secrets or route regulated enterprise queries over unauthenticated gateways. Previously, teams either relied on AgentCore Gateway with MCP tool bridges or paid external search SaaS providers that complicated data-governance reviews in banking, healthcare, and public-sector deployments.
Why MicroVM Isolation Is Winning the Agent Runtime Battle
As enterprises move from chat prototypes to autonomous software that executes shell commands, edits repositories, and queries internal databases, shared multi-tenant containers have proven too risky. Security teams increasingly demand hardware-level isolation per user session, yet traditional virtual machines were historically too slow and costly to spin up per API request.
By applying the same Firecracker snapshot mechanics behind AWS Lambda to stateful, 8-vCPU agent sandboxes with 32 GB of memory, AWS is making dedicated microVM isolation fast enough for interactive user-facing apps. The ability to mount Amazon S3 buckets directly inside the isolated session trust boundary further removes fragile pre-signing and local file-staging scripts, allowing long-running research or financial analysis agents to stream large datasets over internal AWS networking without leaving temporary copies behind on persistent disk volumes.
For engineering teams in Singapore and globally already standardised on Amazon S3 and IAM, folding live web search, S3 mounts, and sub-two-second agent sandboxes into one native control plane significantly lowers the operational cost of shipping production agents ahead of AWS re:Invent later this month. Customers deploying existing V1 runtimes can migrate incrementally by updating their runtime platformVersion configuration to V2 while keeping existing container images and Bedrock model bindings intact.
Sources
- 1AWS What's New: Next-Gen AgentCore Runtime Defaultaws.amazon.com
- 2AWS Machine Learning Blog: Web Search API on AgentCoreaws.amazon.com