Context is King #6 – Building AI Agents on Real Data
Context is King is back for a sixth edition, this time in Paris, kicking off the AI Engineer Paris week. After five sold-out events in San Francisco, London and the Nordics, we're continuing what makes the event work: real implementation stories from builders at the AI frontier.
The focus is building AI agents on real data: what information an agent has access to, how it uses it, and how you ground it in production. Context windows, tools, memory, retrieval, MCP, agent harnesses. The playbook is being written in real time.
We're bringing engineers building with these systems to break it down. Expect implementation stories and live demos.
🎤 Talks & demos
Agentic GraphRAG: why context engineering bottoms out in a graph
Anthony Alcaraz — GTM Agentic Engineering Lead, AWS
Most production agent failures come from context, not from the model. Flat vector retrieval hands an agent a bag of chunks with no structure to reason over, no memory that survives the session, and no way to know what it already knows. This talk makes the case that the substrate for context engineering is a graph, and walks the layers that make it hold up in production: knowledge architecture and memory, reasoning and planning over typed structure, tool orchestration, evaluation and self-evolution, and the efficiency work that keeps the whole thing affordable at scale. Drawn from Anthony's O'Reilly book on agentic GraphRAG.
Graph Engineering: the missing layer between agents and system intelligence
Jeremy Jouvance — AI Engineer & Innovation Tech Lead, Amundi
Prompt and context engineering help models make better decisions. Harness and loop engineering turn those models into capable individual agents. But many real-world tasks require something more: heterogeneous expertise, interdependent work, parallel execution, independent verification, and persistent state. This talk introduces graph engineering as the missing layer between individual agents and system intelligence: how explicit and evolving graphs can organize tasks, coordinate agents, and manage runtime state to turn a collection of capable components into a coherent system that pursues complex objectives reliably.
How much context each check actually needs
Jean Duquenne — CTO, Certo
Certo runs agents on cosmetics regulatory compliance. Different checks need very different amounts of context. Some need an agent searching a full document set. Some need three fields extracted first and a short prompt. Some need one row from a database and no model at all. Give a check more context than it needs and it gets slower and less accurate. Give it less and you get a confident wrong answer. Jean takes three checks running in production and shows what context each one gets, and why, with a live demo of one check given the wrong amount of context.
GoCanopy's harness building journey
Yash Pabbisetti — Co-Founder & CTO, GoCanopy
GoCanopy's journey building an AI analyst for real-estate investment teams: text-to-SQL, then a LangGraph workflow, then a ReAct loop, each stage compounding on the last. How workflows, tools, and agent loops each earn their place, with a live demo on real deal data.
Beyond the agent demo: engineering the production harness
Moustapha Ebnou — Co-founder & CTO, Cominty
An impressive agent demo is not a production system. This talk explores the engineering behind the agentic harness: the execution layer that connects models, tools, context and runtime controls. Drawing on his work at Cominty, Moustapha discusses practical choices around state management, sandboxed execution, failure recovery and observability, and a framework for deciding what to build yourself, what to delegate to managed infrastructure, and how to move beyond the prototype.
Context Gates: automating context building at the source
Stan Dmitriev — AI Product Director, Aiven
Every agent is rich in context for exactly one session. At the moment it creates, updates, or deletes something, it knows the spec, the plan, and the reason better than anyone ever will again. Then the session closes, and all of it dies with the chat window. Stan's fix: context gates. MCP tool calls that block actions unless the agent explains itself, capturing the why at the only moment it truly exists. The system stores and versions that history on the asset itself, so the next agent, or the next human, inherits it instead of guessing. With a live demo.
📅 Agenda
Doors open at 18:30
19:00 – 20:30: Presentations + Q&A
20:30 – 21:30: Networking
👋 Behind the event
Aiven is the AI-ready open source data platform. Get production-ready Apache Kafka, PostgreSQL, ClickHouse, OpenSearch, and more — on any cloud, in minutes, with no infra ops required.
Elaia is a European venture capital firm backing B2B tech and deep tech founders from inception through growth, with a portfolio that includes Vibe.co, Mirakl, Alice & Bob and Proxima Fusion.
Daphni is a Paris-based venture capital firm backing European tech founders from seed to Series A, with a portfolio that includes Back Market, Swile and Pasqal.