Agent runtimes, the protocol layer above them, and retrieval that holds up in production.
This is the domain I'm deepest in right now. Not because I've read the papers, but because I've had to make the unglamorous decisions: where the isolation boundary sits, what a skill is allowed to do, when retrieval is worth the complexity, and when a rule beats a model.
| Technology | Depth | What I did with it |
|---|---|---|
| MCP (Model Context Protocol) | Expert | Shipped a plugin platform and skill repository on it; built a personal server. |
| Agent sandbox architecture | Expert | Isolation model, runtime boundaries, and blast-radius design. |
| Agentic orchestration | Expert | Hybrid rule-engine plus model routing for operations workflows. |
| Agent context / knowledge base | Practiced | Structured markdown as retrievable, self-contained context. |
| RAG architecture | Practiced | Chunking strategy, retrieval quality, and governance guardrails. |
| Multi-agent design | Practiced | Specialist routing and escalation paths. |
| Claude Agent SDK | Practiced | Agent tooling and skill authoring. |
The category's real bottleneck isn't model quality any more — it's everything around the model. Isolation, permissions, retrieval quality, evaluation, cost control. Products that treat those as infrastructure concerns to be solved later are the ones that stall at pilot.