Article
When half of the subway ads promote AI tools for engineers, you know the world is changing. Our team went to the AI Engineer New York 2026 conference to compare notes with the people building at the frontier, and came back with five themes that tell one story: stop waiting for a better AI model and start getting results from the models you have today.
Five themes worth acting on
Theme 1: Capability overhang
Almost no organisation uses AI to its full capability. Even the most advanced frontier labs were not using AI across their whole company until they built proper tools around the models.
There is a huge gap between what we can achieve with models and what we actually do with them. This is called capability overhang. The gap exists mostly because of a lack of tooling and a lack of skill in using the models to their full potential.
Theme 2: The focused AI-native organisation
When your company is AI-native, your agents prompt you more than you prompt your agents. Agents start work on their own or in response to a signal, rather than waiting for a human to tell them to act.
You steer your business by steering your agents, not the other way around. For that, your agents need access to your company’s knowledge.
Theme 3: Agents as trustworthy employees
Agents can now book appointments, buy products, and guide people forward. Trust and judgement have become the limiting resources when working with agents.
An action can be fully legitimate at one moment, and the same action with the same credentials 30 seconds later is not. This leads to intent-based authorisation, which requires treating agents as employees rather than software. Agents need an identity, a role, a budget and a boss, so they do the right task at the right time with the right credentials.
Theme 4: Evals as a quality check for agents
The more autonomous the agent, the more you need a reliable way to check its work. This was the hottest topic of the conference. We heard talks on everything from using chess as an evaluation to using statistics on a golden dataset for judgement-based evaluations.
The uncomfortable part is one where LLM does the work, and another LLM grades it. Neither of the LLMs is deterministic, so the evaluation carries noise, no matter how good the prompts are.
Theme 5: Tracing to keep agents accountable
The more autonomy agents have, the more we need to track what they are doing and when. “Computer says no” is not an answer anyone can defend.
Organisations using agents need transparent tools that record everything and store it long enough for replays, while ensuring that no personal data or anything else you do not want ends up in there. Tracing also helps you further develop your agents. From the traces, you can spot repeated extra steps and eliminate them.
The model is not the bottleneck, the organisation is
All these themes point in the same direction. The gap between what AI can do and what companies get out of it is not closed by waiting for a better model. It is closed by giving agents the right tools and context, organising work around them, treating them with the same expectations as new employees, and measuring and tracing their work so you can trust the results.
At Nitor, we believe this is engineering work. The building blocks are available to your company today.
Written by
