P11 · Agentic AI & AI Systems
IR & RAG, agents/MCP/A2A — systems engineering on top of GenAI
AI as a system: retrieval and RAG grounding LLMs in your documents, and agents that plan, call tools and act — the engineering layer on top of generative models.
Ask a language model about your company's refund policy and it will answer fluently, confidently, and from nothing — it has never seen your policy. Give it tools and let it act on that answer, and the fluency becomes a liability rather than a party trick. This pillar is the engineering layer that turns a generative model into a system: one that consults your documents before it speaks, and whose actions are bounded by something other than its own judgement.
Retrieval comes first, because grounding is the cheaper and more reliable half. The track treats search as search rather than as a preamble to generation: indexing, ranking, and the metrics that tell you whether the right document came back at all — which matters because a retrieval-augmented system with poor retrieval is just a slower way to be wrong. Only then does generation join it, with the model answering from what was retrieved instead of from memory.
Agents follow, and they are where autonomy starts paying and costing at the same time. Tool calling, MCP, and multi-step workflows let a model plan and act; the track keeps the failure modes in view throughout, because an agent that loops, misreads a tool result, or takes a plausible wrong action does so at machine speed and without hesitating. This is the systems half of the LLM story — the architecture itself is P10's subject, and the adversarial view of the same surface is P59's.