18-self-improving-agents

Self-Improving Agents with Strands

Go from a 20-line Strands agent to a fully autonomous, self-improving agent deployed on Amazon Bedrock AgentCore Runtime. Across six progressive steps, the agent learns to create its own tools, rewrite its own system prompt, remember across sessions, work autonomously in the background, and finally deploy itself to production.

Adapted from the AWS New York Summit 2026 builders session AIM308 — "Using Strands to build fully autonomous, self-improving AI agents." Origin: re:Invent 2025 AIM426.

Tutorial Details

Information	Details
Strands Features	Agent, @tool, load_tools_from_directory, dynamic system prompt, ambient background loop
Agent Pattern	Single self-improving agent (extends itself step by step)
Native Tools	shell, file_write
Custom Tools	system_prompt (self-modification), calculator (example of a self-written tool)
Memory	Amazon Bedrock AgentCore Memory (semantic, per-actor)
Complexity	Beginner → Advanced (progressive)
Model Provider	Amazon Bedrock — Claude Opus 4 (global.anthropic.claude-opus-4-8)
SDK Used	Strands Agents SDK, bedrock-agentcore

The core pattern: Agent → Model → Tools

Every step reuses this one loop. Once you grok Agent · Model · Tools, everything else is a variation.

The 6-step journey

#	Step	Capability you add	New code	Runtime	Visual
0	step-00-hello	Baseline agent (one model, one tool)	20 lines	Local	▶
1	step-01-tools	Self-extending — hot-reload ./tools/*.py	+10	Local	▶
2	step-02-system-prompt	Self-modifying — rewrites its own prompt	+40	Local	▶
3	step-03-memory	Self-learning — AgentCore Memory	+30	Local	▶
4	step-04-ambient	Autonomous — ambient + scheduler	+50	Local	▶
5	step-05-deploy	Deployed — AgentCore Runtime ARN	+15	Cloud ☁	▶

What you'll learn

1. 🛠️ Agents can create their own tools while running — no restart, no redeploy (load_tools_from_directory=True).
2. ✏️ Agents can rewrite their own system prompt — permanently changing their own behavior, persisted to disk.
3. 🧠 Agents can learn from every interaction — using Amazon Bedrock AgentCore Memory.
4. 🌙 Agents can run autonomously in the background — exploring a topic while you're idle, then injecting findings.
5. 🚀 Agents can deploy themselves to production — on Amazon Bedrock AgentCore Runtime.

Prerequisites

• Python 3.10+ and pip
• AWS account with Amazon Bedrock model access enabled for Claude Opus 4 (global.anthropic.claude-opus-4-8) in us-east-1
• AWS credentials configured (aws configure or environment variables)
• For steps 3–5: Amazon Bedrock AgentCore access
• For step 5 deploy: Node.js 20+ (the new @aws/agentcore CLI)

Run any step locally

Each step is self-contained and builds on the previous one:

cd code/step-00-hello
pip install -r requirements.txt
python agent.py "hello, what can you do?"

Then progress through the steps:

cd ../step-01-tools     && pip install -r requirements.txt && python agent.py
cd ../step-02-system-prompt && pip install -r requirements.txt && python agent.py
# ...etc

Steps 3–5: create the memory store first

cd code/step-03-memory
pip install -r requirements.txt
python create_memory.py                       # prints a BEDROCK_AGENTCORE_MEMORY_ID
export BEDROCK_AGENTCORE_MEMORY_ID=<the-id>
python agent.py

Step 5: deploy to AgentCore Runtime

cd code/step-05-deploy
export BEDROCK_AGENTCORE_MEMORY_ID=<your-memory-id>
./deploy.sh                                    # creates + deploys via @aws/agentcore CLI

Then invoke the deployed agent from anywhere:

export AGENT_ARN=arn:aws:bedrock-agentcore:us-east-1:...:runtime/aim308_research_agent-XYZ
python invoke.py "what do you remember about me?"

Notebook walkthroughs

Prefer interactive? Each step has a click-through Jupyter notebook in notebooks/:

Notebook	Step
00_hello_agent.ipynb	Hello agent
01_self_extending.ipynb	Self-extending tools
02_self_modifying.ipynb	Self-modifying prompt
03_memory.ipynb	AgentCore Memory
04_autonomous.ipynb	Ambient autonomy
05_deploy.ipynb	Deploy to AgentCore

See it run

Each step was recorded live against the real Strands runtime + Claude. The animated SVG recordings and replayable asciinema .cast files live in casts/ (replay locally with asciinema play casts/01-self-extending.cast).

Cast	Step	What it shows
00-hello	0	20-line agent; one autonomous shell tool call
01-self-extending	1	Agent writes tools/password_generator.py mid-run, then uses it
02-self-modifying	2	Agent rewrites its own prompt → haiku mode; .prompt persists
03-memory	3	AgentCore Memory recall across sessions
04-ambient	4	Background ambient thread + finding injection
05-deploy	5	agentcore create → add agent → deploy → invoke

Project structure

18-self-improving-agents/
├── README.md
├── code/                          # Commit-by-commit agent source (what you RUN)
│   ├── step-00-hello/             # 20-line baseline
│   ├── step-01-tools/             # + self-extending tools (hot-reload)
│   ├── step-02-system-prompt/     # + self-modification (.prompt persistence)
│   ├── step-03-memory/            # + AgentCore Memory (semantic recall)
│   ├── step-04-ambient/           # + autonomous background work + scheduler
│   └── step-05-deploy/            # + AgentCore Runtime deployment
├── notebooks/                     # Interactive click-through mirror of each step
├── images/                        # Animated SVG diagrams (hero + per-step)
└── casts/                         # asciinema recordings (.cast + animated .svg)

Additional Resources

• Strands Agents SDK
• Strands Agents Tools
• Strands Documentation
• Amazon Bedrock AgentCore
• @aws/agentcore CLI

Disclaimer

This sample is provided for educational and demonstration purposes only. It is not intended for production use without further development, testing, and hardening. For production deployments, implement appropriate content filtering and safety measures, follow security best practices, conduct thorough testing, and review configurations for your specific requirements.