Updated October 25, 2026
· ~2 min read
An agent framework decides how you write the loop (model calls, tool calls, state, hand-offs between agents) and a runtime decides where that loop runs safely. In 2026 the two have separated cleanly. Frameworks compete on programming model; runtimes like Amazon Bedrock AgentCore, generally available since October 2025, run agents built with any of them, with session isolation, executions of up to eight hours, identity, memory, a gateway that turns APIs and Lambda functions into MCP tools, and OpenTelemetry-compatible observability.
So pick a framework for how your team wants to write agents, and a runtime for how you need to operate them.
Frameworks leave the production problems to you: isolating one user's session from another's, holding credentials to act on a user's behalf, giving agents tools without hand-writing integrations, and seeing what an agent actually did. AgentCore packages those: Runtime (isolated sessions, up to eight-hour executions, Agent-to-Agent protocol), Gateway (APIs, Lambda functions and existing MCP servers as tools, with IAM or OAuth), Identity (OAuth vaulting and acting on behalf of users), Memory, Observability (CloudWatch dashboards, OpenTelemetry to Langfuse, LangSmith, Arize Phoenix, Datadog and others), and Policy, which checks every tool call against rules before it runs. In June 2026 AWS added AgentCore harness, which stands up a production agent with two API calls.
Question 1 of 2
Is the agent a structured workflow with branches, approvals and resumable state?
Use LangGraph when the agent is a genuine workflow — branches, loops, human approvals, long-running state you need to inspect and resume — because explicit graphs and checkpointing make that tractable. Use the Claude Agent SDK when the agent is mostly a capable model with tools, especially MCP tools, and you want the same harness that powers Claude Code. Use the OpenAI Agents SDK for the quickest path to tool use and hand-offs between agents. Use Strands Agents on AWS for a model-driven loop with first-class Bedrock support and AWS's own maintenance. Use Google ADK if you are building on Google's stack. Then run whichever you chose on AgentCore Runtime, Lambda or ECS — the framework does not lock you to a runtime.
Production GenAI Platform on AWS
Deploy secure, scalable AI systems using AWS-native GenAI services. RAG, agent orchestration, and guardrails: production-ready.
LLM Observability & Evaluation on AWS
See what your AI features actually do: every prompt, response, cost and latency traced, and evaluations that tell you whether a change made things better or worse. Running in your own AWS account.
Secure LLM Gateway on AWS
One governed way for every team to call AI models: keys held centrally, budgets and rate limits per team, fallbacks between providers, and hardening that assumes the gateway is a target.
We work with startups and growing companies on AWS architecture decisions. Tell us about your situation and we will recommend the right approach.
As of October 25, 2026. Provided as is, for general information, and accurate to the best of our knowledge on the date shown. Not legal, compliance, tax, financial or other professional advice. Prices, limits and vendor services change; check current vendor documentation before relying on anything here. Terms of use.