Agent Engineering Bootcamp: Developers Edition

Outcome: Engineer ReAct agents with structured tracing, decision auditing, memory, and termination control.

Framework: Implement perceive-think-act loops, planning agents, reflection, and tool-use architectures.

Hands-on: Building a production-grade agent harness and unpacking what makes Claude Code powerful beyond the model.

Outcome: Serve production LLMs with predictable latency, lower cost, and benchmarks you can defend.

Deep dive: Master GPTQ, GGUF, QLoRA, AWQ, prefill, decode, KV caching, and speculative decoding.

Infrastructure: Measure TTFT, inter-token latency, throughput, and the economics of each deployment choice.

Outcome: Build retrieval systems that route, reason, adapt, and move beyond naive top-k search.

Framework: Implement retrieval-as-a-tool, query planning, multi-hop reasoning, and semantic chunking.

Advanced: Add Graph RAG and semantic caching to improve reasoning quality, cost, and latency.

Outcome: Engineer real-time voice agents with sub-second responses and natural conversational flow.

Architecture: Build streaming STT, LLM, and TTS loops with turn-taking, end detection, and barge-in.

Tools: Compare Deepgram, ElevenLabs, OpenAI Realtime, Vapi, and Retell across latency, quality, and cost.

Outcome: Design multi-agent systems with clear topology, communication protocols, and operational judgment.

Patterns: Implement orchestrator-worker, hierarchical, debate, and handoff workflows for real use cases.

Protocols: Connect agents with MCP and A2A, and evaluate when single-agent systems are the better choice.

Outcome: Ship agent systems with measurable quality, layered safety, and production-ready eval loops.

Guardrails: Implement Llama Guard, NeMo Guardrails, validation, and prompt-injection defenses.

Evals: Build trajectory-based evals with golden tasks, partial-credit grading, and LLM-as-judge scoring.