Road to NODES 2026 | Agents in the Wild: Building AI That Learns

This workshop is designed to push graph-based agent memory beyond its original proof of concept, addressing two questions that remained open after the NODES AI session: does the architecture hold when an agent is truly autonomous rather than following a pre-defined workflow, and does the learning it captures actually generalise into durable, transferable reasoning patterns? You will explore both questions through value investing as a concrete, demanding domain - Agents answering complex questions about companies and their competitive landscape, including ownership structures, peer comparisons, financial ratios, and market positioning. The workshop includes hands-on exercises across four stages, with real code, graph schema designs, Cypher queries, and live demos throughout. You will compare the same analytical task implemented as a predefined workflow versus a fully autonomous agent, examining how their execution traces differ structurally and what each can learn. You will build the full memory stack in Neo4j — hierarchical context graphs, separate retrieval and answer playbooks, and dynamic playbook composition at query time -  and go in depth on the Reflector→Curator loop, where human feedback and external signals like long-term company performance both feed into playbook updates. You will also see how routing specific reasoning steps to specialised models improves output quality, and how that routing logic itself becomes a learned playbook entry. By the end of the session, you will understand how to design a Neo4j schema that captures agentic execution traces at the right granularity, how predefined and autonomous agents interact differently with graph-based memory, and how to build a multi-signal learning loop that improves from both outcomes and corrections. You will leave with a clear view of where graph-based memory pays off, and an honest account of its trade-offs in latency and complexity.