awsai mlPractitioner

Runtime Instances: Persistent Compute for AI Agents on Amazon Bedrock

5 min read AWS BlogAug 6, 2026Reviewed for accuracy

Practitioner — Hands-on experience recommended

In the world of AI, managing workloads efficiently is crucial. Runtime instances in Amazon Bedrock AgentCore provide a solution by offering persistent, managed infrastructure tailored for complex agent workloads. This means your AI agents can operate seamlessly without the overhead of constantly spinning up new environments, allowing for faster and more reliable processing.

At the core of this functionality are microVMs, which create a fully managed environment for invocations. These microVMs can run for up to 8 hours and support stateful workflows through managed session storage. The orchestrator agent on the runtime microVM coordinates tasks, routing them to specialized worker agents that perform compute-intensive operations, such as code compilation or security scanning. You can configure these instances with specific parameters, like the operating system (default is Linux 64-bit ARM), allowed EC2 instance types (default is c7g.2xlarge), and the programming language runtime (default is Python 3.13). This flexibility helps you tailor the environment to your specific needs.

In production, it’s essential to understand how session persistence works. Sessions can last up to 14 days, which is beneficial for long-running tasks. However, keep in mind that while these microVMs are powerful, they may not be suitable for every scenario. Always evaluate your workload requirements and test thoroughly to ensure optimal performance.

Key takeaways

→Leverage Runtime instances for persistent compute in AI workflows.
→Utilize microVMs for stateful workflows with managed session storage.
→Configure parameters like operating system and instance types to fit your needs.
→Understand session persistence of up to 14 days for long-running tasks.

Why it matters

In production, efficient resource management can significantly reduce costs and improve response times for AI agents. Runtime instances streamline operations, allowing teams to focus on development rather than infrastructure management.

Code examples

Python

1writer = Agent(
2    model="us.anthropic.claude-sonnet-4-5-20250929-v1:0",
3    system_prompt=(
4        "You are a senior Python engineer. "
5        "Given a task, return ONLY a single Python code block — no prose."
6    ),
7)
8
9@app.entrypoint
10def handler(event, context):
11    task = event.get("task") or event.get("prompt")
12    session_id = getattr(context, "session_id", None) or event.get("session_id")
13    session_dir = SHARED_DIR / session_id
14    session_dir.mkdir(parents=True, exist_ok=True)
15
16    code = str(writer(task))
17    (session_dir / "code.py").write_text(code)
18
19    return {"agent": "writer", "wrote": str(session_dir / "code.py"), "code": code}

Python

1reviewer = Agent(
2    model="us.anthropic.claude-sonnet-4-5-20250929-v1:0",
3    system_prompt=(
4        "You are a strict Python code reviewer. "
5        "Given code, return 3 bullet points: bugs, style, suggestions."
6    ),
7)
8
9@app.entrypoint
10def handler(event, context):
11    session_id = getattr(context, "session_id", None) or event.get("session_id")
12    code_path = SHARED_DIR / session_id / "code.py"
13    code = code_path.read_text()
14    review = str(reviewer(f"Review this code:\n\n{code}"))
15
16    return {"agent": "reviewer", "read": str(code_path), "review": review}