Video summary
How to Build & Deploy an AI Agent with AWS Strands Agents SDK | Devpost Build Session
Main summary
Key takeaways
Overview of the Workshop (Agent Speedrun)
The session is a hands-on guide to ideate, validate, and deploy an AI agent using the AWS Strands Agents SDK (referred to as “Strands agents” / “strands”).
The presenter emphasizes building agents for production-grade reliability and deterministic control, not just demo success.
Key Concepts & Features Covered
1) Agents “you use” vs “agents you build”
- Agents that you use
- Integrated systems like coding assistants/chat tools that already include agent behavior.
- Agents you build (the hackathon goal)
- You build an agentic system using primitives, then host/deploy it so others can use it later.
2) Strands Agents SDK (AWS open-source agent harness)
- What it is
- An open-source agent harness SDK built by AWS.
- Model agnostic / bring your own model
- No built-in model; works with Bedrock models and also with local/open-weight models and other model providers.
- Lightweight + future-proof
- You can swap in new models without redoing the architecture or rewriting system prompts.
- Multi-agent primitives
- Supports stacking multiple agents to create more complex systems.
- Portability
- Can be hosted anywhere, not only on AWS.
3) Bedrock Agent Core Runtime for Deployment (managed runtime)
- What it is
- A managed runtime service for agent deployments.
- Deployment simplicity
- Described as ~five lines of code to deploy.
- Production capabilities provided
- Memory
- Gateway
- Observability
- Identity & permissions
- Payments
- Multi-tenant architecture support
- Security focus
- Highlighted for enterprise-grade security (e.g., VPC/private-link style integration mentioned).
- Compatibility
- Works with Strands and also supports other frameworks (e.g., LangGraph / LlamaIndex / noted in subtitles).
Workshop Modules (Practical Tutorial Flow)
Module 1: Build a Basic Agent Loop with Tools
Teaches the “reasoning loop” concept:
- User asks a question
- LLM reasons about the task
- Agent selects and calls the appropriate tool
- Tool returns data
- LLM uses tool output to produce the final response
Example built: a customer service agent for an e-commerce store
- Example tools:
- lookup customer
- get order history
- process refund
Implementation notes:
- Starting pattern described as a small amount of code to define the agent and register tools.
- Demonstrates tool calling under the hood (e.g., inferring customer ID from context).
- Includes a multi-turn variant (separate
chat.py-like code).
Module 2: Deterministic Control Using “Hooks” (Guardrails)
Hooks are described as deterministic code injected into the agent loop to enforce rules.
Use cases explained:
- Pre-LLM validation
- Refuse invalid requests/users without wasting LLM calls.
- Before tool call
- Rate limiting and stopping runaway behavior if a tool is failing.
- After tool results
- Block unsafe/undesired outputs from being returned.
Demonstrated: rate limiter hook
- Registers a hook that tracks tool calls (example: max calls = 3).
- When the limit is exceeded, the tool call is blocked and the user gets a controlled message.
This is framed as runtime governance: the agent can think, but behavior is constrained.
Module 3: “Skills” and “Steering”
Skills (context management via modular instructions)
Problem addressed: don’t put everything in a giant system prompt.
Skills approach:
- Put detailed workflows/instructions in a separate skills folder (e.g., Markdown files).
- A skills plugin loads only the relevant skill when needed.
Example skills mentioned:
- account/support issues
- order tracking
- refund processing
Demonstration behavior:
- The agent decides to use a skill (e.g., refund) and follows the steps defined there.
Steering (buddy agent / deterministic enforcement of behavior + tone)
Steering goal:
- Correct bad actions or unwanted behavior before/after execution.
Mechanism:
- A buddy agent or deterministic code steps in to approve/guide/reject actions.
Steering can enforce:
- workflow correctness
- e.g., must look up customer and check order history before refund
- response tone and formatting constraints
- e.g., “don’t use M dashes”
- don’t overpromise/blame the user
- acknowledge frustration
Demonstrated refund flow:
- Checks required prerequisites (customer verified/order history retrieved).
- Evaluates response and tone; only proceeds if checks pass.
Also mentions: session management / session persistence
- Without it, the agent may forget earlier conversation context.
- Deployment to Agent Core includes session handling support.
Workshop Logistics & Learning Resources
- Uses AWS Workshop Studio
- Fully free, available for ~72 hours
- Provides a Jupyter notebook environment and workshop materials
- Access notes
- Avoid logging in via an existing Amazon account.
- Use an email that won’t associate with a billed AWS account.
- Provides
- a code editor URL to open notebooks
- a GitHub repository and workshop materials for later reference (screenshot/repo access emphasized)
- Presenter guidance
- Only module 1 is sufficient to submit for the hackathon.
- Additional modules improve determinism and reliability.
Main Speakers / Sources
- Sandy — Senior Developer Advocate at AWS (primary presenter)
- Darlles / Dar Developer — Devpost Developer Marketing Manager (host/introducer)