tl;dr. aws bedrock agentcore is a managed runtime for production ai agents on aws — framework-agnostic, model-agnostic, priced per invocation, with seven primitives (runtime, memory, gateway, identity, browser, code interpreter, observability) that handle the scaffolding so you write only the agent.
what is aws bedrock agentcore?
aws bedrock agentcore is a managed runtime for production ai agents on aws. it is framework-agnostic (bring strands, langgraph, the vercel ai sdk, or your own loop) and model-agnostic (point it at bedrock, anthropic, openai, or anything else), priced per invocation, with seven primitives — runtime, memory, gateway, identity, browser, code interpreter, observability — that own the scaffolding so you write only the agent.
in concrete terms, agentcore is a set of managed services for the parts of an ai agent that are annoying to build yourself: a runtime that scales, memory that persists, a way to turn apis into tools, and identity that isn't a custom iam nightmare. pricing is per-invocation for runtime and per-operation for the supporting services, so an idle agent costs nothing and a busy one scales until you put a ceiling on it. regional availability is commercial aws — us-east-1 and us-west-2 lead, others follow on the usual aws cadence. sessions are sandboxed per-invocation, cold starts sit in the 3–8s range, and events stream back over sse or the sdk's async-generator protocol.
what agentcore is not: a framework for writing the agent logic itself. you still write that. agentcore is what runs underneath.
the seven primitives
there are seven services. you don't need all of them. most production agents touch three or four.
- runtime serverless container for your agent. you hand it an entrypoint function, it scales, sandboxes each session, and streams events back.
- memory short-term (session, multi-turn) and long-term (persistent across sessions). memory is the thing you most regret not wiring up.
- gateway wraps apis, lambdas, and other services as mcp tools the agent can call. fewer custom integrations, more schema-driven tool definitions.
- identity inbound (who called the agent?) and outbound (what can the agent reach?). supports cognito, okta, auth0, entra id, and the obvious oauth flows to third parties.
- code interpreter a sandboxed javascript and typescript execution environment. useful when the llm writes code; dangerous if you run code yourself.
- browser a managed headless browser your agent can drive. for the tasks where the cleanest api is still "click the button a human would click."
- observability opentelemetry traces of the whole agent lifecycle — tool calls, memory reads, llm turns, errors — piped to cloudwatch.
the shape of a minimal agent
here is the smallest useful program. the real sdk exposes
BedrockAgentCoreApp from bedrock-agentcore/runtime;
you pass it an invocationHandler with a
process async generator, and each yielded event becomes a
server-sent event. in production you call app.run() to boot
a fastify
server on port 8080; in this sandbox we invoke the handler directly so
you can see the events without leaving the browser.
read the trace. app.init fires when the app object is created,
app.entrypoint_registered when the handler is wired up,
app.yield for each event the generator emits, and
app.done when the invocation closes. that sequence is the same
shape you'll see in cloudwatch once you deploy.
try it: change the prompt, add another yielded event, return a plain object instead of yielding. the mock doesn't care — it mirrors the real sdk's permissive handler contract (async generator for streaming, plain async function for one-shot).
what this course is and isn't
this is eight lessons for engineers who want a working mental model of agentcore without spending an afternoon on iam policies first. every lesson ships with a runnable example. nothing you type leaves your browser.
lesson eight is the handoff to real aws — costs, deploy, and what changes when the mock goes away. everything before that is about understanding the shape of the thing, so that when you do deploy, the deploy is boring.