AIToday
Large Language ModelsAI Safety & AlignmentITmedia AI+Published: Nov 6, 2026, 06:01 JST

AGNTCon + MCPCon Japan 2026: Harness, not model, decides AI agent success

Send on LINE
AGNTCon + MCPCon Japan 2026: Harness, not model, decides AI agent success

3 Key Points

  1. What happened

    At AGNTCon + MCPCon Japan 2026, Mazin Gilbert said open-weight stacks cost roughly one-sixth as much, yet 80% of companies still choose frontier models, and Angie Jones used a $48 refund that became $4,800 to argue teams, not models, own the outcome.

  2. Why it matters

    The same model can produce completely different results depending on its harness, so production reliability depends on what teams build around the model — sandboxes, permissions, retries and oversight — not on model choice alone.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The event, held September 10–11, 2026 in Shibuya, Tokyo, was the first Japan edition of AGNTCon + MCPCon. Gilbert built his case from Artificial Analysis rankings, where frontier models such as Claude and GPT lead but open-weight models like GLM and Kimi are rapidly closing the gap. He pointed to the Jevons paradox: token prices fell more than 90% over the past year while consumption rose roughly 50-fold or more, so total spending grew even as unit costs dropped. Jones's exercise drew its force from the audience vote, in which the team received the most votes as the responsible party. Her conclusion was that only humans can bear responsibility, making own the loop — the ability to verify execution logs afterward, constrain agent actions, and change the system when needed — a prerequisite for production deployment.

FAQ
What caused the $4,800 refund in the example?
Order data correctly stored 4800 cents, but the model dropped the 'cents' unit when calling a tool, so 4800 was processed as dollars.
How much cheaper are open-weight model stacks according to Mazin Gilbert?
He said building a stack with open-weight models costs about one-sixth as much, yet 80% of companies still choose frontier models.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleSOX semiconductor index sinks 11% as AI sell-off hits Celestica