$2M+ · AI Infrastructure

What they're building

Jetty is building infrastructure for agentic AI workflows across three primitives that solve for the same underlying problem: agents that ship reliably, not just run. The core concept is the runbook: a plain markdown file that gives an AI agent its job, its definition of done, and a method for checking its own work before calling it finished. Skills (the instructions) plus standards (the definition of done) equals a runbook. That equation is the product thesis in five words. An agent running against a runbook does not just execute a prompt. It generates output, verifies that output against defined criteria, retries if anything fails, and escalates to a human when it cannot resolve a failure after a fixed number of attempts. The agent definition lives as a markdown file in your repo, not as a graph in a framework or a config inside a vendor's console. You own it. Any model runs it. On top of runbooks, Jetty provides sandboxed execution environments (isolated workspaces where agents run without touching production systems), real-time tracing from instruction to output (every decision in the chain is logged), and evaluation loops that ingest traces, run assessments, and produce pull requests with verified improvements. Jonathan's Substack post "Generation Got Cheap. Verification Didn't." is the longest and most rigorous public articulation of the problem: as token costs collapse 30-80%, teams automate more tasks without expanding their verification capacity. The fraction of output that is actually verified shrinks with every new task the model takes on. Jetty is the verification infrastructure that closes that gap.

Why this matters

Jonathan's framing for the market timing is precise and worth understanding. He cites a 2026 MIT paper by Catalini, Hui, and Wu that formalizes what practitioners already feel: the cost to generate AI output is collapsing, the cost to verify it is not. The authors call the expanding zone between those two curves the Measurability Gap: the growing share of tasks where machines can cheaply generate output that humans cannot affordably verify. Jetty sits inside that gap as the tool that makes verification scalable. This is not a niche problem. Gartner projects 40% of enterprise applications integrating task-specific agents by the end of 2026. IDC projects AI copilots embedded in 80% of enterprise workplace applications in the same timeframe. Every one of those deployments faces the same question: how do you know the agent is working? How do you catch a regression when you swap models? How do you prove compliance to a regulator when the agent made a decision that affected a patient or a financial account? Current answers are: manual spot-checking, vibe-based dashboards, and crossing fingers. Jetty's runbook-plus-evaluation infrastructure is the systematic alternative. The Mila affiliation is also meaningful for the market. Mila is one of the two or three most productive AI research institutes in the world: Yoshua Bengio, one of the three people who won the 2018 Turing Award for deep learning, is its scientific director. Being embedded in the Mila Ventures space puts Jetty at the centre of a research-to-deployment pipeline that no other city in Canada has.