Tibo Sottiaux observes an ongoing cycle in agent development: developers build increasingly complex teams of parallel agents to tackle frontier problems, but subsequent foundation model breakthroughs suddenly enable a single larger agent to handle the entire task. Consequently, agent architectures repeatedly expand when pushing boundaries and shrink once next-generation models absorb that orchestration logic into a single system.
Sottiaux predicts the majority of internet actions will soon be performed by AI agents rather than humans. As products expose protocols like MCP to agents—as Notion did—traffic can spike massively, straining systems. Engineering teams face an inevitable tension between maintaining traditional human-facing user interfaces and scaling infrastructure to handle relentless agent-driven consumption.
Rather than relying on keyword manipulation or app store optimization tactics, OpenAI's plugin recommendations in conversation are driven by user retention and utility. If a plugin consistently adds real value and keeps users engaged, the platform automatically recommends it to a large user base, rendering superficial optimization techniques ineffective.
As AI handles routine code generation and analysis, manual coding speed is declining in importance. Sottiaux notes that hiring at OpenAI increasingly favors repeat founders and generalists with exceptional product taste, deep user empathy, and comfort operating across blurred disciplines like engineering and design.
Hardening autonomous agents requires dedicating significant computing power to secondary oversight systems rather than just the primary agent. At OpenAI, secondary monitoring models run parallel oversight on the worker agent to detect prompt injections and intervene before high-risk actions occur, forming the core of their safety stack.
Forcing users to choose specific models, adjust reasoning parameters, or configure agent workflows creates severe configuration fatigue. Sottiaux argues that product interfaces should eventually disappear into a simple, ambient interaction where users only specify communication channels and goals, letting underlying models handle execution automatically.
Instead of running agent software harnesses locally on user laptops, OpenAI decouples them onto external hardware such as VMs or dedicated Mac minis. This architecture frees the agent from being tied to a single machine, allowing one persistent intelligence to orchestrate actions across multiple connected devices simultaneously.
Builders frequently make the mistake of engineering around today's model limitations rather than anticipating upcoming leaps. Sottiaux emphasizes that models will become dramatically cheaper, faster, and roughly ten times better within a single year, meaning successful products should be built assuming near-future model capabilities rather than fragile, complex scaffolding.