I'm Eric McLaren. By day I lead Copilot enablement at Canadian Tire. Outside it I run a production multi-agent platform, live 24/7 since January 2026. Most people describe AI systems. This page just shows you mine.
Research I queued at midnight arrived as a finished document by sunrise. My morning starts at the conclusions, not the search bar.
It watched my inbox and my money and briefed me at 7:03 — so my hours went to the work that needs a human, not the admin that doesn't.
Nine pages of notes became an organized PDF before I parked. Same 24 hours, roughly twice the output.
My PC, driven end to end by one of my agents. It reads the posting, picks the tailored resume it wrote, answers the screening questions, and submits. My part is the judgment call — which role, which pitch. It handles the forty fields.
August 2026. A few seconds of phone video, then the screen capture at 20x speed. The agent picks the tailored resume it wrote, fills every field, answers the screening questions, and it ends on the live confirmation screen.
One self-hosted platform, five specialized agents, running around the clock on my own hardware.
Months of context in a continuously consolidated knowledge base, retrieved by hybrid vector and text search. The agents audit their own recall for errors.
A secured daemon gives agents screen reading, mouse and keyboard control of a Windows workstation. That is what you are watching in the video.
A speech-to-speech stack benchmarked across three vendors and tuned to sub-second first response. It takes live calls.
An evaluation harness catches hallucinated quotes before anything can cite them, alongside detectors that report their own blind spots instead of failing silently.
Overnight research arrives as finished documents. Financial monitoring and scheduled maintenance run on their own clock while I sleep.
Fallback chains and per-lane cost governance across Anthropic, OpenAI, Google and xAI. Inference spend cut roughly 60% at constant quality.
Everyone agrees AI should help people do work. Where it actually dies is adoption: the tool ships and nobody uses it. This loop is how I get from demo to daily habit — it built the platform above, and it ran the Copilot rollout at my day job.
Pick a real, painful, repeated task. Not a demo. Something with stakes and an owner.
Smallest version that touches the real system, on day one, not after a quarter of design.
Every workflow breaks agents in its own way. Each failure becomes a guardrail or a rule.
Never trust the agent's self-report. Check the thing it changed in the real world.
Repeat until the humans stop checking. That's the finish line, and it's measurable.
I co-lead it. Same loop, pointed at an organization: finished things that people actually use — not pilots, not decks, working practice.
Same skill as everything above, pointed at an organization.
Every system on this page is real, self-hosted, and in daily use. I am happy to walk through the architecture, the failures, or the cost model.