Agents Get Started

get-started

Walk a developer from zero to a running agent on AWS.

When to use

• Developer wants to build an agent on AWS and doesn’t know where to start
• Developer wants to create a new AgentCore project
• Developer is choosing between frameworks (Strands, LangGraph, GoogleADK, OpenAI Agents)
• Developer just ran agentcore create and wants to know what to do next

Do NOT use for:

• Environment/prerequisite issues (CLI not found, credentials broken) → use agents-debug
• Adding capabilities to an existing project (memory, tools, policies) → use agents-build or agents-connect
• Migrating an existing Bedrock Agent → use agents-build (loads references/migrate.md (opens in a new tab))

Input

$ARGUMENTS can be:

• A framework preference: “using LangGraph”, “with Strands”
• A protocol: “MCP server”, “A2A”
• A description of what the agent should do: “a customer support agent”
• Empty — the skill will guide framework selection

Process

Step 0: Verify CLI version

agentcore --version

This skill requires v0.9.0 or later.

If the version is older:

Your AgentCore CLI is out of date (found vX.Y.Z, need v0.9.0+).

Offer to run the update: agentcore update. After the update completes, re-check the version to confirm it’s ≥0.9.0 before continuing. Preserve any context the developer already provided (framework preference, project name, what they want to build) so they don’t have to repeat themselves.

If agentcore is not found:

The AgentCore CLI isn’t installed. Run npm install -g @aws/agentcore (requires Node.js 20+). If you’re having trouble with installation, I can run the agents-debug skill (which loads references/doctor.md (opens in a new tab)) to diagnose your environment.

Step 1: Determine intent — exploring or ready to create?

Before jumping into framework selection, figure out where the developer is:

Ask the developer: “Are you exploring options (comparing frameworks, understanding what AgentCore does) or ready to create a project?”

• Exploring → Go to Step 2 (framework comparison). Present the options, answer questions, and wait. Do not construct a create command until they signal they’re ready.
• Ready to create → Skip to Step 3 (create the project). If they already specified a framework, skip Step 2 entirely.
• Already has a project → Look for agentcore/agentcore.json in the current directory. If found, read it and skip to Step 5 (what to do next). Don’t re-scaffold.

If the developer’s intent is clear from $ARGUMENTS (e.g., “create a Strands agent called MyBot”), skip straight to Step 3.

Step 2: Framework selection

Check conversation context first. If the developer already discussed frameworks earlier in this conversation (e.g., from a previous skill invocation), don’t re-present the full table. Summarize what was discussed and ask if they’ve decided, or if anything changed.

If this is the first time discussing frameworks, present the options:

Supported frameworks (CLI-scaffolded, Python):

Ask the developer to choose. Present the options and wait for their selection. Don’t assume a default unless they explicitly say they have no preference.

Note on naming: The CLI flag value is the exact string to pass to --framework. In prose use the shorter names.

Default recommendation (only when the developer says “no preference” or “you pick”): Strands — AWS-native framework with the tightest AgentCore integration and the most samples/docs.

Key decision points to surface:

• “Do you have existing agent code in LangGraph or OpenAI Agents?” → use that framework
• “Do you need complex graph-based workflows with conditional branching?” → LangGraph
• “Starting fresh with no preference?” → Strands

Framework not listed?

If the developer asks about a framework not in the table above, handle it:

BYO Container path (any framework, any language):

For frameworks or languages not scaffolded by the CLI, AgentCore Runtime accepts any container that implements the HTTP contract (POST /invocations, GET /ping). The workflow:

1. agentcore create --name <ProjectName> --defaults to scaffold the project structure
2. agentcore add agent --type byo --build Container --language <Language> --code-location <path> to register your code
3. Write a Dockerfile that builds and runs your agent
4. agentcore deploy handles ECR push, CDK infra, and runtime creation

Language-specific notes:

The rest of this skill (deploy, status, logs, invoke) applies once the container builds correctly.

Framework vs. model provider — a common confusion

The framework is how your agent orchestrates (Strands, LangGraph, etc.). The model provider is which LLM it calls (Bedrock, Anthropic, OpenAI, Gemini). These are independent choices:

• Strands + Bedrock (default) — AWS-native everything
• Strands + Anthropic — Strands orchestration, direct Anthropic API for the model
• LangGraph + Bedrock — LangGraph orchestration, Bedrock for the model
• OpenAI Agents + OpenAI — OpenAI everything

If the developer says “I want to use Claude” they mean the model provider (Bedrock or Anthropic), not the framework. If they say “I want to use LangGraph” they mean the framework.

Step 3: Create the project

Build the agentcore create command based on the developer’s choices.

Before constructing the command — validate the project name. The CLI fails late: if the name is invalid, you’ll see the error after walking through prompts or building the full command. Save the round-trip and check these rules up front. Reject the name and ask for a new one if any rule fails:

• Length ≤ 23 characters (this is shorter than most developers assume — MyCustomerSupportAgent is 22 chars and fits; CustomerSupportChatbot is 22 and fits; MyCustomerSupportBotApp is 23 and just fits; MyCustomerSupportChatBot is 24 and fails)
• Alphanumeric only — no hyphens, underscores, dots, or spaces
• Must start with a letter

Say the count back out loud when close to the limit: “That name is 24 characters — the CLI caps project names at 23. Want to shorten it to <suggestion>?” Do not run the command with an invalid name on the assumption that the CLI error message will be clear — it isn’t always, and the developer’s mental model will be wrong for subsequent commands.

Construct the command, then present it for confirmation before the developer runs it. Show the full command with all flags and explain what each choice means. Wait for the developer to confirm or adjust before proceeding.

Example presentation:

Here’s the command I’d recommend based on what you’ve told me:

bashCopy

agentcore create --name MyAgent --framework Strands --model-provider Bedrock --build CodeZip --memory none

This creates a Strands agent using Bedrock models, deployed as a code zip (no Docker needed). Memory can be added later.

Want to run this, or change anything?

Do NOT execute the command automatically — present it and wait.

Minimal (defaults — Strands, Bedrock, CodeZip, no memory):

agentcore create --name <ProjectName> --defaults

With specific options:

agentcore create \
  --name <ProjectName> \
  --framework <Framework> \
  --model-provider Bedrock \
  --build CodeZip \
  --memory none

Flag reference:

Guidance on choices:

• Protocol: Use HTTP unless the developer specifically needs MCP tool serving or A2A agent-to-agent communication
• Build: Use CodeZip unless the developer needs custom system dependencies (CodeZip is faster to deploy and doesn’t require Docker locally)
• Model provider: Use Bedrock unless the developer has a specific reason for another provider (Bedrock doesn’t require managing API keys)
• Memory: Start with none — memory can be added later via agents-build (loads references/memory.md (opens in a new tab)) when the developer needs it

Step 4: Explain what was created

After the project exists, read agentcore/agentcore.json and the generated code to explain the project structure.

The layout below reflects CLI v0.9.x. If the CLI version is different, run tree <ProjectName>/ -L 3 to see the actual generated structure and explain from there.

<ProjectName>/
├── agentcore/
│   ├── agentcore.json      ← Project config (agents, resources)
│   ├── aws-targets.json    ← AWS account + region
│   ├── .env.local          ← Local environment variables (gitignored)
│   └── cdk/                ← CDK infrastructure (auto-managed, don't edit)
└── app/
    └── <AgentName>/
        ├── main.py          ← Your agent code — this is where you build
        ├── mcp_client/      ← Pre-wired example MCP client (see note below)
        └── pyproject.toml   ← Python dependencies

Key files to highlight:

• app/<AgentName>/main.py — the agent’s entry point. This is where the developer adds tools, system prompts, and logic.
• agentcore/agentcore.json — the project config. Resources are added here via agentcore add commands.
• agentcore/.env.local — local environment variables. After deploy, resource IDs are written here for local dev.

Heads-up on the scaffolded MCP client. main.py imports get_streamable_http_mcp_client() from mcp_client/client.py and appends it to tools. In a fresh project, this client points at a public example MCP endpoint — so agentcore dev works immediately. Two things to flag:

1. It will become a silent no-op if you repoint it at a gateway that isn’t deployed yet. The common path is to swap the example endpoint for os.getenv("AGENTCORE_GATEWAY_<NAME>_URL"). That env var is only populated after agentcore deploy. If the developer repoints and runs agentcore dev before deploying, get_streamable_http_mcp_client() returns a client with a None URL and the agent starts with zero MCP tools — no error, no warning. See the “Local dev gap” section in agents-connect for the guard pattern: if not GATEWAY_URL: tools = [].
2. If the developer doesn’t need MCP tools at all, remove the mcp_clients list and the loop that appends it to tools. The scaffold includes it as a convenience, not a requirement.

The reference client code in agents-connect (Path A) shows the correct pattern for gateway-backed MCP clients once deploy has run.

Step 5: Local development

agentcore dev

This starts a local dev server. The developer can interact with their agent immediately.

Port the dev server binds to (important if you’re scripting curl calls or testing from another process):

The CLI prints the bound port and URL on startup — always read the actual value from the CLI output rather than hardcoding. If the default port is already in use, the CLI auto-increments (e.g., 8080 → 8081 → 8082), so a second dev session or a lingering process from a previous run can shift your port without warning. Use agentcore dev --port <N> to pin it, or grep ps / check the CLI banner if invocations start failing with connection-refused or exit-code-7 errors.

Important limitations to mention:

• Memory is not available in agentcore dev — it requires a deploy
• Gateway URLs are not available locally — they require a deploy
• The local server uses the model provider configured in the project

Step 6: First deploy

When the developer is ready to deploy:

agentcore deploy

This will:

1. Show a preview of AWS resources to be created
2. Ask for confirmation
3. Build and deploy via CDK

First deploy takes 3-5 minutes. Subsequent deploys are faster.

After deploy, show them how to invoke:

agentcore invoke "Hello, what can you do?"

And how to check status:

agentcore status

Step 7: What’s next

Based on what the developer said they want to build, suggest the logical next skill:

Don’t overwhelm — suggest one or two next steps based on what the developer actually asked for.

Example walkthroughs

For task-framed prompts (e.g., “build a customer support agent”), load the matching example reference:

More examples can be added to this skill’s references directory as common patterns emerge.

Output

• A clear path from “I want to build an agent” to a running deployed agent
• The agentcore create command tailored to their choices
• An explanation of the generated project structure
• Concrete next steps based on their intent

Quality criteria

• The agentcore create command uses only valid flags from CLI v0.9.1
• Framework recommendation is based on the developer’s context, not a generic default
• The developer understands what each generated file does
• Next steps are specific to what the developer wants to build, not a generic list of all features