Agentic Development Series

0.1 Series Overview

I went through Andrew Ng’s Agentic AI course on DeepLearning.AI mainly because I kept hearing “agentic” thrown around at work without a shared definition, and wanted a rigorous one before building anything else on top of Hermes Agent. The course is good, but every example is generic — customer support bots, research assistants. None of it touches what I actually do all day: segmenting voids and lifted chips out of X-ray scans. So I rebuilt every example from scratch, in my own domain, following the course’s own five-module order — then kept going for four more posts covering things the course doesn’t touch at all but that I think you need if you actually want to run this in production.

If you read all nine posts and work through the code, you should be able to go from zero to building, hardening, and shipping your own agent.

0.1.1 Posts in this series

The course’s own five modules, rebuilt around segmentation:

1. What Agentic AI Actually Is — the four design patterns (reflection, tool use, planning, multi-agent), the domain rules that make this a real problem instead of a toy one, and setting up aisuite.
2. Reflection — teaching an agent to catch a physically impossible segmentation mask (a void bigger than the chip it’s in) before it reaches a KPI report, both self-critique and multimodal external critique.
3. Tool Use and MCP — the biggest post in the series: aisuite’s automatic tool-calling, the manual tool-calling loop, and a full MCP server so Copilot, Cursor, and Claude Code can call the same tools directly.
4. Evaluations and Error Analysis — the module Andrew Ng himself calls the most valuable: the component-tagging method for turning a vague “the model isn’t great” into a ranked, concrete list of what to fix next.
5. Planning and Multi-Agent Systems — JSON plans vs. code-as-plan, and the two ends of the multi-agent spectrum: a predictable linear pipeline and a chaos-tolerant all-to-all team.

Four posts past the course, covering what it takes to actually ship this:

6. Advanced Agentic AI Development — retries, timeouts, cost tracking, memory, prompt-injection defenses, automated testing, Docker/systemd deployment, OpenTelemetry tracing, a real HITL queue, and prompt/config versioning. The longest post in the series.
7. Agent-to-Agent Interoperability with A2A — Google’s A2A protocol for when independent agents, built by different teams in different frameworks, need to talk to each other as peers, not as a tool call.
8. Fine-Tuning and Distilling Your Own Agent — logging real production traffic, filtering out human-corrected examples, and LoRA fine-tuning a small local model to reproduce a large model’s agent decisions cheaply, with a staged rollout.
9. Retrieval, Grounding Your Agent in Real Knowledge — two different retrieval problems (visual case similarity vs. textual spec knowledge), plus why whole-image similarity and defect-crop similarity need separate embedding spaces to be useful.

0.2 Tools used throughout

• aisuite as the model-provider layer — one interface, so the same agent code runs against a cloud model or a local Qwen3 endpoint on my own GX10 with a one-line change
• MCP (Model Context Protocol) and A2A (Agent2Agent) for, respectively, agent-to-tool and agent-to-agent interoperability
• DINOv3 + LanceDB for visual case retrieval, sentence-transformers for textual spec retrieval
• Real semiconductor-imaging domain rules throughout: p2/p98 percentile X-ray normalization, chip-containment checks, IoU-based evals

Note

This series pairs well with the Local-AI Defect Inspection Series, which covers fine-tuning and evaluating small local VLMs on the same class of imagery.