sample-bedrock-agentcore-agent-skill

A community sample Agent Skill (packaged as an Agent Plugins v1 plugin) that helps you build AI agents on Amazon Bedrock and Amazon Bedrock AgentCore. It is a decision router: point your agent (or yourself) at it and it maps a use case to the right stack and the right reference file.

This is a community sample and not an official AWS distribution. It is not released, endorsed, or supported by AWS. Content was authored clean-room from official primary sources; see PROVENANCE.md.

What's inside

plugin.json                         # Agent Plugins v1 manifest (validates against the 1.0.0 schema)
skills/bedrock-agentcore-agents/
  SKILL.md                          # the router: decision tree + progressive disclosure
  references/                       # one file per domain, loaded on demand
  assets/                           # decision-tree diagram, Terraform + IAM templates, agent.py, checklist
PROVENANCE.md                       # file -> source-URL ledger + clean-room method
LICENSE                             # MIT-0 (MIT No Attribution)

The skill under skills/bedrock-agentcore-agents/ is self-contained: copy that folder on its own and it still works.

Domains covered

Model choice + Bedrock Converse API · tools with Strands @tool / MCP · RAG with Bedrock Knowledge Bases · multi-agent patterns · serverless production on AgentCore Runtime · AgentCore Memory · Gateway + Identity · Guardrails · observability · testing & rollout · Terraform-first IaC · security/IAM/cost · and when a managed alternative (Bedrock Agents, Flows, Responses API) fits better than code-first.

Install & use

This is a standard Agent Skill, so it works in any client that supports the Agent Skills standard. Install it once, then just describe what you want to build — the client loads the skill when your request matches its description (progressive disclosure: metadata first, then SKILL.md, then references on demand).

git clone <this-repo>
git clone https://github.com/ThejashreeNeralakere/sample-bedrock-agentcore-agent-skill.git
cd sample-bedrock-agentcore-agent-skill

### What to copy

**Copy exactly one folder: `skills/bedrock-agentcore-agents/`**

It is self-contained — all internal links are relative and one level deep — so it works on its own
with no dependency on anything else in this repo. Copy the **whole folder**, not individual files:

skills/bedrock-agentcore-agents/ <-- COPY THIS ENTIRE FOLDER ├── SKILL.md required the router; the client reads this first ├── references/ (19 files) required loaded on demand by SKILL.md └── assets/ (9 files) optional templates you copy when building ├── decision-tree.md ├── service-selection-matrix.md ├── model-selection-guide.md ├── agent/agent.py ├── iam/bedrock-invoke-least-privilege.json ├── terraform/{guardrail,knowledge-base,agentcore-runtime-role}/ └── checklists/production-readiness.md

`SKILL.md` alone will load, but it routes to `references/` — so dropping those loses the actual
content. Keep the folder intact.

**Do not copy these into your skills directory** (they are repo-level, not part of the skill):

| Not copied | Why |
| --- | --- |
| `plugin.json` | Agent Plugins manifest — **only** used when installing as a Kiro Power (see below), where you point at the repo root instead of copying |
| `PROVENANCE.md`, `LICENSE`, `README.md`, `CONTRIBUTING.md`, `CODE_OF_CONDUCT.md` | Repo governance/provenance; the skill declares `license: MIT-0` in its own frontmatter so licensing travels with it |
| `tests/`, `.kiro/specs/` | Repo validation and authoring history |

**One exception — Kiro Power:** you don't copy anything. You point Kiro at the **repo root** (the
folder containing `plugin.json`) and it reads `plugin.json` + `skills/` itself.

### Kiro

Kiro supports both the plugin wrapper (a **Power**) and plain skills
([Powers](https://kiro.dev/docs/powers/), [Skills](https://kiro.dev/docs/skills)).

**As a Power** (keyword activation + one-click updates):
1. Open the **Powers** panel (Ghosty icon with the lightning bolt)
2. **Add Custom Power** → **Import power from a folder** (or from this repo's Git URL)
3. Select the repo root (the folder containing `plugin.json`) → **Install**

**As a plain skill:**
```bash
cp -r skills/bedrock-agentcore-agents ~/.kiro/skills/     # all workspaces
cp -r skills/bedrock-agentcore-agents .kiro/skills/       # this workspace only

Also available as the /bedrock-agentcore-agents slash command. Manage under Agent Steering & Skills.

Claude Code

Skills live in .claude/skills/ (docs):

cp -r skills/bedrock-agentcore-agents ~/.claude/skills/   # personal, all projects
cp -r skills/bedrock-agentcore-agents .claude/skills/     # project-scoped

Codex

Codex reads skills from ~/.agents/skills (global) or .agents/skills (repo) (docs):

mkdir -p ~/.agents/skills
cp -r skills/bedrock-agentcore-agents ~/.agents/skills/   # global
cp -r skills/bedrock-agentcore-agents .agents/skills/     # repo-scoped

Codex uses the same progressive disclosure model and can select the skill implicitly from its description, or you can invoke it explicitly.

GitHub Copilot

Copilot looks in .github/skills, .claude/skills, or .agents/skills (project) and ~/.copilot/skills or ~/.agents/skills (personal) (docs):

cp -r skills/bedrock-agentcore-agents ~/.copilot/skills/  # personal
cp -r skills/bedrock-agentcore-agents .github/skills/     # project

Agent skills work with the Copilot cloud agent, Copilot code review, the Copilot CLI, the Copilot app, and agent mode in VS Code.

Tip — one location for Codex and Copilot: both read ~/.agents/skills, so a single copy there covers both.

Verify the copy

After copying, confirm the skill landed intact (adjust the path for your client):

SKILL_DIR=~/.claude/skills/bedrock-agentcore-agents   # or ~/.kiro/skills/... , ~/.agents/skills/... , ~/.copilot/skills/...

test -f "$SKILL_DIR/SKILL.md" && echo "SKILL.md: ok" || echo "SKILL.md: MISSING"
echo "references: $(ls "$SKILL_DIR"/references/*.md 2>/dev/null | wc -l | tr -d ' ')  (expect 19)"
echo "assets:     $(find "$SKILL_DIR/assets" -type f 2>/dev/null | wc -l | tr -d ' ')  (expect 9)"

The folder name must stay bedrock-agentcore-agents — it has to match the name in SKILL.md frontmatter, or clients will reject the skill.

Then prompt naturally

You don't open reference files by hand. Try:

• "Build a customer-support agent on AWS that answers from our PDFs and remembers each customer across sessions."
• "Productionize my Strands agent on AgentCore Runtime with guardrails, observability, and Terraform."
• "Which built-in evaluators should I use to catch hallucinations?"
• "Write least-privilege IAM for a Bedrock agent."

How to tell the skill loaded. The answer should cite official AWS/Strands URLs, lead with the managed-vs-code-first decision, prefer the Converse API over InvokeModel, mention the AgentCore Runtime ARM64 + /invocations + /ping container contract, and flag Preview features instead of defaulting to them. Generic advice with no citations means it didn't load.

Reading it directly (no client)

1. Open skills/bedrock-agentcore-agents/SKILL.md
2. Follow the decision tree to the reference file(s) for your use case
3. Use the assets/ templates as starting points when you build

Important: volatile facts

This sample never hard-codes prices, quotas, current model IDs, or GA/Preview status. Where those matter, it links to the live source (AWS console, model cards, pricing, quotas). Always confirm time-sensitive details at the source before relying on them. Anything labeled Preview should not be used as a production default.

Provenance & licensing

• Authored only from official primary sources (AWS docs/blogs, Strands docs, official AWS GitHub orgs, the Terraform Registry, and package indexes). Every recommendation carries an inline citation; the full mapping is in PROVENANCE.md.
• Licensed under MIT-0 (MIT No Attribution) — see LICENSE.

Contributing

See CONTRIBUTING.md. Note: contributing this sample to aws-samples requires confirming with AWS Open Source Program Office (OSPO) and signing the Contributor License Agreement (CLA) first — this has not been done yet.