InfoQ roundtable debates autonomy limits for enterprise agents

On November 3, 2026, InfoQ reported on a roundtable discussion hosted by its Geek Yuge Yue live program, produced jointly with QCon. The topic was where to draw the line on autonomy for enterprise-grade agents. The discussion was moderated by Wang Fang, CTO of NoDesk AI, and featured Lü Detao, algorithm specialist at Fliggy; Qi Longji, technical director at lalamove; and Wang Shangqi, CTO for AI at Jiangsu Mobile.

The central question the panel addressed is what happens when an agent moves beyond answering a question and begins placing orders, participating in scheduling, modifying production systems, or calling real business capabilities. In that world, a single mistake is no longer a matter of user experience but of actual operational risk. The panelists framed the core challenge as balancing model uncertainty against business determinism, which in turn requires a clearly defined boundary that keeps the model's unpredictability contained within a deterministic operational frame.

On specifics, the panelists recommended that critical workflow steps be governed by rules, with AI used only to describe and draft those rules, so that the system remains explainable, reversible, and controllable. Operations touching user wallets or financial entitlements must be kept on the human side as a safeguard, while tasks with more latitude, such as composing replies, clarifying ambiguous intent, or handling long-tail phrasing, can safely be delegated to the model. They also argued that an enterprise's real edge in building agents does not come from producing answers that merely look good, but from whether the company's own data, toolchains, algorithms, business rules, and evaluation systems can hold up and be validated end-to-end in real deployments.

The report also notes that AICon, InfoQ's global AI development and application conference, will take place in Beijing on December 18–19, featuring thirteen technical tracks spanning post-training, inference optimization, agent engineering, and knowledge engineering.