Persistent Bedrock Runtime Instances let multi‑agent workflows run across days

Small teams can stop fighting serverless session time limits: Bedrock AgentCore Runtime Instances keep agent workflows and shared state running across days, cutting rework and coordination overhead.

What actually changed

On 30 September 2026, AWS added a second compute option for Amazon Bedrock AgentCore: Runtime Instances. Where Bedrock previously offered MicroVMs as a serverless host for short-lived agents, Runtime Instances provide AWS‑managed EC2 infrastructure for persistent, long‑running agent workflows. The blog demonstrates a three‑agent music pipeline: one agent runs a generative audio model on the instance’s own GPU; two other agents open the .wav file it writes from a shared volume. The post also walks through creating capacity providers, deploying agents from different artifact types, orchestrating agent‑to‑agent collaboration using shared sessions, and persisting workflows across multiple days.

Who this affects

• Small teams building creative or stateful multi‑agent workflows where work stretches beyond short serverless sessions (for example, music production, long editing passes or multi‑step data curation).
• Operations or infra owners deciding whether to use serverless MicroVMs or persistent compute for agents: MicroVMs are for short, isolated sessions; Runtime Instances are for workflows that must keep GPU compute, files and session state available over days.

What it costs or what it replaces

• MicroVMs remain the serverless option. The blog highlights their benefits: fast cold starts, session isolation and consumption‑based pricing, and uses the phrase "session isolation" when describing that model.
• Runtime Instances replace the need to fragment long workflows into many short sessions by providing AWS‑managed EC2 instances that hold GPU and filesystem state for agents across days.
• The announcement does not state pricing. The post contrasts consumption‑based MicroVM billing with a persistent Instance model but gives no numbers or instance types; you will need to check AWS account pricing pages or contact your sales rep for concrete costs.

What we don't know

• Exact pricing and billing model for Runtime Instances compared with MicroVMs.
• Supported GPU models, instance sizes and region availability for Runtime Instances.
• Operational limits (maximum session duration, concurrency caps, scaling behaviour) for Runtime Instances and capacity providers.
• Security and isolation differences between MicroVMs and Runtime Instances beyond the high‑level descriptions.
• Compatibility edge cases for agent artifact types not shown in the music example.

What to do next

1. Pilot this week: pick a small, stateful workflow you already run (for example, a one‑track music test or a multi‑step media transcode). Day 1: create a capacity provider and provision a single Runtime Instance in a non‑production AWS account. Deploy one agent that runs a generative audio model and verify it can write a .wav to a shared volume.

1. Day 2: deploy two additional agents from different artifact types (as the blog describes) that open the .wav file from the shared volume and operate via the same session. Confirm that session state and files persist overnight and that agents can resume work the next day.

1. Day 3: compare behaviour to an equivalent MicroVM run — measure cold starts, session continuity, and operational friction. Record costs from your billing console and check whether Runtime Instances reduce manual orchestration or reprocessing. If costs are unclear, contact AWS sales or consult your account pricing pages before moving to production.

What to do next

1. If the pilot shows value, draft a runbook that specifies when teams should choose MicroVMs versus Runtime Instances based on session length, GPU needs and persistence requirements.
2. Schedule a short budget review with finance using actual billing from the pilot; do not proceed to production until instance sizing and pricing are confirmed.
3. If security or limits matter, open a support ticket to clarify isolation, quotas and region availability before wide rollout.