Building<with>Agents

Most teams building self improving agents follow a similar trajectory: the agent runs, reflects, updates an artifact, and the metric rises. In production, the metric can still rise while a later incident traces back to a learned change that passed review but was evaluated against the wrong signals.

This talk examines different self learning systems that may change the agent context, the agent harness, or model weights. Each offers different gains and failure modes. Memory can absorb sycophancy and inflate its reward signal. Skill libraries can overfit to the last task or be tainted by one bad run. Harness changes can discard useful behaviour as new tools arrive. Weight updates persist longest and are the hardest to reverse.

I’ll cover the controls that make self iteration viable in production: bounded, reversible updates; provenance for every learned artifact; promotion gates; and evaluations that separate improvement from noise.

To turn the art of building AI agents into a science, you need a way to measure progress — and measuring an agent turns out to be far harder, and far more important, than measuring a model. This talk is about how to actually evaluate agents: what to measure, how the field gets it wrong, and why good measurement is what eventually lets agents improve themselves.

What we’ll explore:

• What you’re really testing when you evaluate an agent — the model vs. the harness around it
• The building blocks of an eval: traces, datasets, and scoring — and the dimensions most people miss
• Where today’s benchmarks and LLM-as-judge break down, with real examples
• A practical playbook for measuring agent quality in development and in production
• How measuring success well opens the door to self-improving, self-healing agents

How do you move from supervising every agent action to trusting a system that works in the background? This talk uses feedback control systems as a lens for designing agent architectures that can take on ongoing work with increasing autonomy. We’ll explore the core primitives: an environment with tools and memory, clear objectives, observable outcomes, verification, and a controller that determines when the agent should act, pause, or ask for help.

Examples such as optimizing ad performance, improving e-commerce margins, and reducing application latency will show how different jobs call for different feedback cycles and architectural choices. We’ll examine how budgets, tool permissions, and limits on work awaiting review shape autonomy, alongside natural ways for people to guide and steer agents through feedback. A live demonstration will bring these concepts together, starting with an environment and an agent and developing a repeatable system for doing a job, reviewing its results, and incorporating feedback. Attendees will leave with a practical framework for choosing suitable work and expanding agent autonomy as trust grows.

Most GenAI talks measure success by speed. We think that misses the bigger opportunity and hides a risk. An AI that accepts its inputs at face value will confidently repeat upstream mistakes. An AI that one practitioner uses well rarely makes the next practitioner any better.

At BMO, we built a reusable AI skill that turns scattered model-development material into final model documentation, starting with models that support the BMO Blue Rewards loyalty program. It did make documentation faster and more consistent. It also did something we didn’t design for: when it pieced the whole model story back together, it surfaced gaps and inconsistencies that earlier development and review had missed. We also saw the same skill produce very different documentation depending on who used it and how hard they questioned it. Both patterns match published research: LLM-generated model cards are more complete than human-written ones, LLMs rarely catch errors reliably on their own, and AI assistants spread the behaviour of top performers to everyone else.

We call the practice that follows Skills Engineering. Prompt engineering tunes a single interaction. Skills Engineering designs how AI takes part in a workflow: what evidence it demands, which checks it hands to deterministic tools, when it escalates to a human, and how the lessons from each use feed back into the skill for the next practitioner. We will share the design patterns, show a skill critiquing public model cards live, and be honest about where the approach fails.

Enterprise AI work has shifted from “generate text” to “make a decision”: routing, scoring, classification, triage. New models built to skip autoregression entirely (a category sometimes called “System 1” AI) promise order-of-magnitude speed and cost gains for exactly these tasks, without giving up transformer-grade reasoning. For teams running AI in real-time or high-volume pipelines, this is a shift worth understanding.

Enterprise AI work has shifted from “generate text” to “make a decision”: routing, scoring, classification, triage. New models built to skip autoregression entirely (a category sometimes called “System 1” AI) promise order-of-magnitude speed and cost gains for exactly these tasks, without giving up transformer-grade reasoning. For teams running AI in real-time or high-volume pipelines, this is a shift worth understanding.