Has AWS opened the back-end of AgentCore Runtime?
The rollout of Lambda microVMs has prompted some to draw a connection. It must be said that agent workloads appear to be the main target of this offering.
Here you find the Firecracker virtualization technology, paired with SnapStart (using snapshots to avoid cold starts). However, where traditional Lambda functions are aimed at event-driven applications and have a maximum execution time of 15 minutes*, the microVMs can run for up to eight hours and do not require a function manager (the method that handles events). They also provide specific capabilities (shell access, mounting file systems, installing packages…), by default based on an Amazon Linux base image.
You package the application code with a Dockerfile that Lambda executes. Once the environment is initialized, it captures a snapshot that is then used to launch the microVMs. Each one has its own public endpoint with JWE authentication, handling HTTP/1.1, HTTP/2, WebSockets, gRPC and SSE.
AWS offers five base configurations ranging from 0.5 GB of memory and 0.25 vCPU to 8 GB and 4 vCPU. Each can scale up to 4x, billed by the second. They can be suspended during periods of inactivity — with memory and disk state preserved — and automatically resumed when traffic returns. During this time, you only pay for the storage of the snapshot.
A Reference Deployment for Claude Managed Agents
Lambda microVMs target multi-tenant applications where each user needs a dedicated environment to run third-party code. Among them are vulnerability scanners, data analytics platforms, CI/CD tools and game servers that run scripts provided by players. They also include sandboxes for AI.
AWS provides a reference deployment for Claude Managed Agents. The orchestration (the model and agent harness) remains with Anthropic, while the tooling runs inside microVMs (one per Claude session). The only inbound traffic is the webhook call. The rest of the workflow runs on a pull basis.
ClickHouse Integrates Lambda MicroVMs with its chDB Engine
In addition to the environment variables defined at image creation, you can pass configuration parameters to a given VM at runtime.
The Lambda microVMs are currently available in a handful of AWS regions, including one in Europe (Ireland), on ARM64. ClickHouse is among the companies that have already implemented them to support various use cases for its in-process analytic engine, chDB. Among others:
- Federated queries joining local data, S3, CDN and Postgres
- Isolated CI/CD executors
- On-demand sandbox to reproduce bugs
- Local memory for each agent session
* Excluding ‘durable functions’, whose execution can span up to a year, reinforced with checkpoints and replays. An option introduced in 2025.
