What it's for
DIY cloud is a process and a dotenv. AgentCore is isolation, identity, a wrap, a rule, memory that outlives the process.
Jobs
Runtime
The LangGraph you ran on a laptop now has to survive a support spike.
A process on a box. Cold start is yours. Two users share memory if you are not careful. Scale is a ticket.
Runtime takes the same graph. Dedicated microVM per session. Fast cold start. You did not rewrite the loop.
Isolation
Maya and Jordan both ask about refunds. They must not see each other's files.
One process, two threads, a shared /tmp. A write in A is a bug report in B.
Session A is a microVM. Session B is another. End A, it is sanitized. B never saw the file.
Identity
The agent should post the refund note to Slack as Maya, not as a shared bot.
A workspace token in a dotenv. Every session is the bot. Consent is a comment in the prompt.
Outbound OAuth. No token, the Gateway tool is blocked. Maya granted Slack; the post is hers.
Gateway
Refunds and order lookup already exist. Every new agent wants them.
Copy the REST client into each repo. Drift. Three ways to refund $80.
Wrap the API and the Lambda once. One Gateway endpoint. MCP tools for any agent.
Policy
The model asks to refund $800. Policy is $500.
A sentence in the system prompt. Sometimes the model obeys.
Cedar or natural language on Gateway. Intercept every tool call. DENY before the API. Deterministic.
Memory
Maya said keep it short. She comes back in four weeks.
A new process. Thursday never happened. Or you stuffed the transcript into a prompt.
Short-term died with the session. Long-term still has tone=terse. Not a LangGraph checkpointer.
Patterns
Use it when you need production: isolation, an IdP, an MCP wrap, a deterministic policy, durable memory, traces. Skip it if you are still on localhost with a single user and a dotenv.
Two clocks. The LangGraph checkpointer is the cursor through this run. AgentCore Memory is the conversation (short) and what should still be true next month (long).
| LangGraph checkpointer | AgentCore Memory | |
|---|---|---|
| Scope | This run / this thread | Short-term: this session. Long-term: this user, next month |
| Holds | Snapshots you can replay | The conversation, then preferences and insights |
| Reach for it | Pause, resume, time travel inside one graph | A new session should still know they like terse replies |
FAQ
No. It is an agentic platform. Any framework, any model. LangGraph is the loop. AgentCore is how that loop runs in production: isolation, identity, tools, policy, memory that outlives a process.
If the graph still lives on localhost with one user and a dotenv, no. Reach for AgentCore when two customers share a host, when an IdP has to say who may call, when the company API should be wrapped once, when $500 is a rule, when terse replies must still be true next month, or when you need to find a denied refund in a trace.
Harness is a config loop: model, prompt, tools, isolated microVM, no orchestration code. Runtime deploys the LangGraph, Strands, CrewAI, or custom agent you already wrote. Same factory floor.