Amazon Web Services announced the general availability of its next-generation Amazon Bedrock AgentCore runtime and a dedicated serverless Web Search API on November 11–12, 2026, overhauling how autonomous AI agents execute code and retrieve grounded information. The release introduces Firecracker microVM snapshot restoration that cuts session cold starts from roughly thirty seconds to under two seconds at P75, alongside elastic scale-to-zero memory reclamation and direct AWS Identity and Access Management authentication for real-time web retrieval.
Next-Generation Bedrock AgentCore Runtime Architecture
The updated AgentCore runtime (Platform Version V2) re-architects serverless execution for autonomous agents by assigning every session an isolated Firecracker microVM. Unlike conventional multi-tenant function environments that share underlying kernels across invocations, each AgentCore microVM enforces strict hardware-level isolation for compute, memory, and ephemeral storage. All session data is destroyed when the agent task completes, preventing lateral contamination between sequential or concurrent executions.
To eliminate the latency penalty historically associated with booting complex containerized environments, the runtime captures memory snapshots after container initialization. Subsequent agent sessions restore directly from this pre-warmed snapshot image, reducing P75 startup latency to under two seconds. Workloads can allocate up to 8 vCPUs and 32 GB of memory per session—a fourfold increase over the initial generation—and mount custom container images up to 10 GB. This compute envelope allows agents to perform resource-intensive operations within the isolated boundary, such as compiling software libraries, running headless browser automations, or executing local model inference.
Memory efficiency is managed elastically through automated reclamation and scale-to-zero capabilities. Idle agent sessions release allocated memory within minutes of inactivity, reducing billed memory by up to 68% for interactive or multi-turn conversational agents. Direct Amazon S3 FUSE mounts allow agent sessions to access large external datasets securely without copying files or exposing static credentials into the runtime filesystem.
Dedicated Web Search API on the AgentCore Data Plane
Alongside the runtime overhaul, AWS launched the Web Search API via the InvokeWebSearch operation on the bedrock-agentcore data plane. The API connects autonomous agents directly to an Amazon-maintained web index without requiring third-party search subscriptions, external API tokens, or custom gateway middleware.
Calls to InvokeWebSearch return structured search results comprising page titles, URLs, ranked text extracts, and publication timestamps when available. Developers can apply granular filters by publication date ranges and specific domain lists to constrain retrieval to verified repositories. Because requests use standard AWS Signature Version 4 (SigV4) authorization, permissions are governed entirely through AWS Identity and Access Management policies targeting the managed tool resource arn:aws:bedrock-agentcore:[Region]:aws:tool/web-search.v1. All retrieval queries can be routed privately over AWS PrivateLink and audited through AWS CloudTrail data events.
The Web Search API is available under pay-as-you-go pricing at $7 per 1,000 queries in six launch regions: US East (N. Virginia, Ohio), US West (Oregon), Europe (Ireland), and Asia Pacific (Tokyo, Sydney).
Distinction from Local Emulators and Prior Tooling
The cloud-native AgentCore V2 runtime operates distinctly from local developer tooling such as the AWS emulator released earlier in the month. While local emulators simulate cloud endpoints on developer laptops for offline testing and rapid feedback, the next-generation AgentCore runtime executes on managed AWS cloud infrastructure designed for multi-tenant production agents. Research teams building agentic workflows must distinguish between testing orchestration logic on local emulators and deploying production agents into isolated Firecracker microVMs with enforced CloudTrail auditing and IAM boundary protections.
Implications for Academic Labs and Research Offices
The convergence of microVM sandboxing and audited web retrieval carries significant operational implications for higher education and institutional research administration:
- Sandboxed Untrusted Code Execution: Autonomous agents increasingly generate and execute Python code, simulation scripts, and statistical pipelines. Running untrusted code within a dedicated Firecracker microVM isolates potentially dangerous processes from university networks, ensuring that flawed agent scripts cannot compromise institutional infrastructure.
- Institutional Compliance and Research Security: Because the Web Search API and AgentCore runtime utilize native IAM authorization, VPC routing, and CloudTrail auditing, research computing teams can maintain strict governance aligned with institutional Research Security standards and federal data protection protocols.
- Grounded Literature Retrieval and Provenance: Live web search capability enables academic agents to support Retrieval-augmented generation workflows that fetch current scholarly preprints and regulatory notices. Constraining queries with domain filters safeguards Training data provenance and verifiable citations.
- Reproducible Agent Workflows: Preserving snapshot configurations, container digests, and search query logs strengthens Computational reproducibility across experimental agent pipelines generating AI and ML research outputs.
- Snapshot State Architecture Caveats: Because the runtime instantiates sessions from pre-boot snapshots, variables initialized during container startup are baked into the snapshot state. Research software engineers must ensure that researcher tokens, participant identifiers, and sensitive research data are handled strictly within dynamic request handlers rather than global initialization blocks.
Sources
- Amazon Web Services. “Run agents securely and efficiently with the next generation AgentCore runtime.” AWS Machine Learning Blog, November 12, 2026. Checked 2026-11-14.
- Amazon Web Services. “Introducing the Web Search API on Amazon Bedrock AgentCore.” AWS Machine Learning Blog, November 11, 2026. Checked 2026-11-14.
- Amazon Web Services. “Amazon Bedrock AgentCore Developer Guide.” AWS Documentation, November 2026. Checked 2026-11-14.
