An event to look beyond the model
The Linux Foundation, through its subsidiary Agentic AI Foundation (AAIF), organized the first AGNTCon + MCPCon Japan in Tokyo’s Shibuya district, held on September 10 and 11, 2026. According to the ITmedia AI+ source, this was the event’s first Japanese edition; the focus of the proceedings was AI agents and the MCP (Model Context Protocol), which some of its main developers discussed.
The account reveals a shift in perspective: over the two days of sessions, discussion did not center only on model capabilities but above all on what lies outside the model—namely, enterprise systems and the ways agents connect to them. It is a sign that the debate is moving from the technology showcase to the conditions for real-world use.
The cost question and the Jevons trap
Mazin Gilbert, executive director of AAIF, argued that models are converging toward similar performance: the Artificial Analysis data cited in the article indicate small differences among the various products. Gilbert reportedly showed his chart before even commenting on it precisely to underscore this uniformity.
Frontier models such as Claude and GPT stand out, rapidly followed by open-weight models such as GLM and Kimi. According to Gilbert, with the latter the cost per token drops to about one-sixth, but about 80% of users would still choose frontier models. The speaker invoked the Jevons paradox, formulated by nineteenth-century economist William Stanley Jevons: if efficiency reduces unit cost, overall consumption can increase. Applied to agents, the mechanism is that tokens per individual operation fall while total consumption grows by about 50 times or more.
Gilbert’s concluding argument, as reported by the source, is that discussing models is not enough: the useful question is whether the entire task can be completed at the lowest cost. One cited figure: the most expensive model costs about $3 per task, while models with performance about ten points lower remain below $0.1. The choice would depend on the task, not on leaderboard supremacy.
The “harness” and the responsibility that remains human
Gilbert shifted attention to the harness, the infrastructure surrounding the model: according to him, even a valid model can produce poor results if the harness is inadequate. It would include the planning and execution loop, tool management, context, memory, the sandbox and error handling, with a perimeter that extends beyond the development process to the entire operational lifecycle. The proposed distinction is sharp: playing with a model in chat is not the same as managing a system in production.
The most concrete case was recounted by Angie Jones, VP of developer experience at AAIF. A customer service agent was supposed to refund a customer an amount equal to $48 but returned $4,800. The stated cause is a formatting error: the record said “4800 cents,” but the model, when calling the tool, reportedly omitted the unit and treated the number as dollars.
Jones invited the audience to ask who was responsible and identified four elements to examine: the team, the runtime, the model, and the tools and data. Her thesis is that responsibility remains with the team, because it is the team that built the system around the model and exposed it in production. Hence the formula “own the loop”: Human in the loop is not enough, nor is Human on the loop, in which the human intervenes from outside. What is needed is traceability of operations, control of agent behavior, and the ability to modify the mechanisms when necessary.
What remains to be verified
The material does not clarify whether the refund case actually happened or is a teaching example, and it does not report which model or vendor was involved. The figures on costs, usage shares, and performance differences come from the speakers’ statements and references cited during the event, not from independent verification. The figures and the generalizability of the case therefore remain to be confirmed, along with the actual scope of the guidance provided during the conference.
