Orchestration d'agents

Mathieu Grymonprez, Global CDO of Adeo, names agent orchestration as the open source brick that is missing: registry, lifecycle, permissions, skills. His verdict on what exists today: « rien en open source ne marche bien pour ça ». What is at stake for him is lock-in of enterprise intelligence, the agentic harness and its "adeo.md", which is why he watches standard Kubernetes, API portability and memory.

In the McKinsey description of an AI operating system, orchestration of reusable agents sits as one layer next to a semantic layer of ontologies and knowledge graphs, deterministic analytical models the agents call, and an experience layer. The same article puts a figure on the distance between deployment and result: 40% of large companies have deployed agents at scale, while 6% attribute at least 5% of their EBIT to AI.

Fred Plais treats orchestration as a human competence rather than a platform. The rare profile he describes knows agent failure modes, chains agents together, and catches an error before it propagates. He considers that profile recent and still scarce, while many teams keep recruiting for code output, which is precisely the resource that stopped being scarce.

The working definition stays narrow: declarative or coded agents, reusable, with routing between LLMs according to the problem. The term does double duty, an architectural layer in one account and an individual skill in another, and in neither form does a usable open source implementation exist.