Building Local and Serverless LLM Agents

🎯 Agenda 🚧 Evolving

⚡ This agenda is a living outline — it is subject to change and will very likely be extended with new topics as the technology and tooling keep evolving. Check back for updates!

🖥️ Local, Zero-Dependency Java Agents (zsmith)

• Building zero-dependency business agents with zsmith
• Single-file agent scripts with no build step
• The agentic loop explained from the source
• Composing tools, profiles, and custom schemas
• Tool permissions and safe execution control
• Skills as portable prompt snippets
• Episodic memory persisted across runs
• Subagents and parallel delegation
• Serving local agents over HTTP
• Local on-device inference
• Provider swapping without code changes
• Built-in agent observability

☕ MicroProfile (Quarkus) + LangChain4j

• Building agents with MicroProfile (Quarkus)
• The tool-calling loop hidden by AI Services
• Composing tools as annotated CDI beans
• Persistent chat memory across sessions
• Streaming responses to the client
• Structured outputs mapped to records
• Guardrails for input and output
• Local inference with Ollama
• Model provider swapping via configuration
• Observability with OpenTelemetry and Micrometer
• Exposing agents over the Agent2Agent (A2A) protocol
• Testing LangChain4j agents

☁️ Serverless on AWS Bedrock AgentCore

• Serverless agents on AWS Bedrock AgentCore Runtime
• The agent loop run by AgentCore Runtime
• Exposing Lambda tools via AgentCore Gateway
• Short-term and long-term AgentCore Memory
• Tool access secured with Cognito and JWT
• MCP-based tool access over the gateway
• Managed cloud inference with Bedrock models
• Foundation model swapping via configuration
• Containerizing zsmith or Quarkus agents with CDK
• Cloud observability with CloudWatch tracing
• From local script to serverless deployment

🔗 References

• bce.design — Boundary, Control, Entity architecture
• airails.dev — AI rails for Java development
• sbce.space — Spec-driven BCE