Amazon Bedrock Launches 14-Day Runtime Instances for AI Agents

Q: If multiple AI agents share a single filesystem, who resolves the inevitable conflicts when they overwrite each other's work?

The developer does. AWS is providing shared infrastructure, not automatic conflict resolution. AgentCore runtime instances let multiple agents share a filesystem and session state for up to 14 days, but nothing in AWS’s announcement suggests Bedrock will detect, merge, or arbitrate file-level write conflicts between agents. In the demo, the pattern was safe because one agent wrote a file and another only read it. That is very different from two agents editing the same artifact at once. In practice, teams have to add the same controls used in distributed systems and shared storage. The relevant techniques are file locking, atomic writes, versioning, append-only event logs, or optimistic concurrency checks that reject stale updates and force a retry. The broader guidance around long-running agents points in the same direction: persist only durable state, use atomic writes, add schema and corruption checks, and log state transitions so failures become debuggable. If developers do nothing, the default outcome is usually “last writer wins,” which can silently overwrite another agent’s work. Some systems preserve both versions by renaming the older file, but Bedrock has not announced that kind of built-in behavior for runtime instances. So the shared filesystem is best seen as a coordination primitive, not a referee. It makes multi-agent collaboration easier, but the application owner still has to define how conflicts are prevented, detected, and resolved.

3 articles · Updated · AWS Blog · Aug 6

Amazon Bedrock AgentCore added runtime instances, a managed EC2-based option that lets production AI agents persist shared sessions for up to 14 days and run continuously across multi-step workflows.
The new compute tier targets workloads that outgrow AgentCore microVMs’ 8-hour limit, especially agents needing GPU acceleration, direct OS access, or multiple collaborating agents on the same host.
Runtime instances let teams deploy multiple agents with separate dependencies in one runtime, share a file system within a session, and stop or restart sessions to cut idle costs while keeping state intact.
Amazon said the service uses the same AgentCore APIs, identity controls and observability as microVMs, with support at launch for Linux ARM64 and x86_64, Python 3.11-3.14, and container images.
The offering is available in 9 regions, including N. Virginia, Oregon, Frankfurt and Tokyo, and is priced at standard EC2 rates plus an AgentCore management fee.