Amazon Bedrock AgentCore Adds Server-side Web Search, Payments, and Long-lived Runtimes

AWS just handed platform teams a new attack surface and called it a feature: AgentCore's recent releases add server-side Web Search, payments, and runtime instances that let agents execute long-running tasks, call external APIs, and persist memory — all inside AWS and all now auditable in ways we haven't had to manage before.

This is not incremental polish. Web Search reached GA as a server-side capability that grounds Bedrock models with current web results fetched inside your AWS environment, so agents can retrieve up-to-date information without embedding ad-hoc external fetch logic in customer workloads. Bedrock has also expanded cross-Region inference for newer model tiers, improving low-latency access from multiple regions.

What materially changes for platform teams is AgentCore itself. The release notes show three linked shifts:

Runtime instances and temporal policies that provide controlled, long-running execution environments and let agents call external APIs for longer-lived workflows.
Payments reached GA, introducing an agent-level billing surface and spend controls tied to agent identity and policies.
New data-ingestion APIs and broader data-plane integrations (S3 objects, EFS access points), private key–based JWT authentication for agent identity, per-agent CloudWatch logs and tracing, and PrivateLink options for secure evaluations.

Taken together, these are the primitives of production-grade autonomous agents: grounding (server-side web search), memory (data ingestion), execution (runtime instances), identity (private key–based auth), observability (per-agent logs/traces), and billing (payments). Better to provide these primitives in a constrained, auditable platform than leave teams to invent credential injection, ad-hoc cron trickery, or opaque sidecar processes.

The security and operations implications are large. Runtime instances create a new trust boundary: agents that can run for minutes or hours and that can mount S3/EFS require identities treated like service accounts with least privilege and lifecycle controls. Payments add a "denial-of-wallet" attack vector, so policy, quotas, and spend alerts must be first-class controls. Private key–based JWTs and PrivateLink reduce credential sprawl but demand tighter key rotation, secrets management, and audits tied to per-agent logs and traces.

The EKS containers blog post about a break-glass pattern for when federated identity fails feels related: operators need auditable, minimal-blast-radius escapes for control-plane recovery. That pattern belongs in your playbook for agent runtimes too — you need secure, auditable escape hatches for misbehaving agent processes or runtime features that outlive their IAM constructs.

Smaller but practical: Lambda now offers managed runtimes in public preview for newer Node.js and Python versions, reducing the need to maintain custom runtime images. For teams that want new language features without custom images, this lowers maintenance burden and should accelerate migrations across serverless fleets.

If you want a concrete next step, treat agents like services. Give them dedicated identities, per-agent log groups and traces, quota'd payment profiles, and playbooks for reclamation and break-glass. Instrument resource and billing telemetry before you let an agent loose.

This wave of features is not about novelty; it's about operationalizing agents. Expect two kinds of teams to win: those that add SRE and security controls around AgentCore now, and those that ignore the new surfaces and get surprised by runaway costs, leaked data, or an agent that persisted credentials into an EFS mount. The only sustainable posture is to assume agents will behave like code you must run, monitor, and, when necessary, cut off.

Amazon Bedrock AgentCore Adds Server-side Web Search, Payments, and Long-lived Runtimes

Sources

Amazon Bedrock adds OpenAI-style APIs and cross-Region inference routing

Amazon EKS now supports multiple external OIDC providers per cluster

Amazon Bedrock AgentCore runtimes GA: AWS-managed GPU EC2 instances for stateful agents and 14-day sessions