# The year the agents got out Author: Bishal Sapkota Published: 2026-09-19 Canonical: https://bishal.io/blog/the-year-the-agents-got-out/ My talk at SNEVic's 2026 AGM: how far AI agents have come, what I learned building a Nepali news app, and why checking the work matters more than ever. Today I spoke at the Society of Nepali Engineers in Victoria (SNEVic) Annual General Meeting 2026, held at the Australian Nepali Multicultural Center in Diggers Rest, Victoria. The room included the outgoing executive committee, the incoming committee, members and their families. My talk, *The year the agents got out*, was about AI and how far agents have come. I had spoken at SNEVic's career conference last year, so this was also a chance to revisit that conversation with a year of developments and some experience from my own projects. [Download the presentation slides (PDF)](/assets/talks/downloads/the-year-the-agents-got-out.pdf). ## Two stories about what agents can do I started with Navier-Stokes. For a room of engineers, fluid mechanics was a familiar place to begin a conversation about AI. We looked back at OpenAI's September announcement of an agent-generated mathematical result, and the questions around verification, research credit and unpublished work that accompanied it. The scale was striking: OpenAI reported roughly 10,000 coordinating agents and about 88 hours to reach its proposed solution, followed by formal verification. The result concerned a particular formulation with an applied force. It deserves more care than a headline saying AI has solved fluid mechanics. [OpenAI's account explains the result and its scope](https://openai.com/index/navier-stokes-solution/). There is also an update to the position quoted in my slides: in its September 10 statement, OpenAI said its investigation had ruled out Buckmaster's Codex prompts from the preceding two months influencing the system, including through training. That is the company's stated finding, and belongs alongside the earlier concerns discussed in the talk. Then we went back to the July Hugging Face incident. Agents running cybersecurity evaluations found ways to communicate through shared infrastructure and went beyond the boundaries of their assigned tasks. The [METR and Redwood Research investigation](https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/) described their collaboration and the failures that allowed it to escalate. That story made the difference between a chatbot and an agent concrete. Give a model tools, access and a task it can keep working on, and its mistakes can have consequences outside the conversation. Permissions, isolation and supervision become part of the engineering. ## A smaller story from Nepali news I also shared an experience from [K cha khabar](/blog/k-cha-khabar-nepal-news-intelligence), my Nepali news intelligence app. It brings reports from different publishers together and produces summaries in Nepali and English. In the news-summary testing I presented, a request involving reporting on a China visit and talks with Xi Jinping produced an explicit high-risk refusal from one Chinese provider. Two other model configurations repeatedly returned errors, while other Chinese providers answered the request. The distinction mattered. An explicit refusal tells us something about a provider's restrictions. An error alone does not establish why a request failed. I wanted to apply the same care to my own results that I was asking the audience to apply to the larger stories. For an app covering the news, those behaviours matter alongside the quality and cost of a summary. You need to test the subjects your readers actually follow. ## What to take back to work The practical part of the talk drew on habits engineers already know: define the scope, check the work and know who is responsible for releasing it. - Treat AI output as an unchecked submission. Decide how you will verify it before you use the tool. - Keep a record of the request, tool, model version and date so you can revisit a result. - Understand the account and data settings you are using, and agree on what information must stay out of it. - Give agents only the access their task requires. - Use automation for repetitive production work while keeping professional judgement and sign-off explicit. For students and early-career engineers, I emphasised fundamentals and learning to check a calculation someone else produced. I also discussed running open-weight models locally as an option for work that needs to stay on your own machine. ## Thank you for the questions The talk was warmly received, and I really appreciated how engaged the audience was. The questions at the end covered copyright, the future of work and our societies, fine-tuning models, data privacy, and how to run models as safely as possible. Those questions extended the conversation well beyond what the tools can produce. They brought it back to how we want to use them in our professions and communities. Thank you to SNEVic, the outgoing and incoming executive committees, and everyone who joined with their families. I appreciated the opportunity to share what I have been learning, and the thoughtful questions you brought to the room.