Most "research" in 2026 is a chatbot guessing and hoping you don't fact-check. That's a gamble with nicer formatting.

When I need to actually know something, I don't search — I deploy. I ask a question, Pi spins up the right subagent: marketing research, project scouting, competitor deep-dives — each one purpose-built. They cross-reference real sources, pull live docs, and compare approaches. And it's not just web scraping. I feed them actual books, scans by experts, deep reference material that never shows up in Google. The agent weighs internet takes against published expertise and comes back with receipts — links to source code, methodology comparisons, concrete trade-offs. Not "here's five blue links, good luck."

Everything lands in my knowledge base, organized by project — each client has its own space, research notes alongside planning docs and decision logs. Pi remembers what we've already figured out, so no question starts from zero. Like a research team that actually reads the memos. Which is more than most human teams manage.

The real shift: I don't dig. I ask better questions. The agents handle the seventeen-sources-and-cross-reference-everything part. I stay in creative mode. If that sounds like dodging grunt work... yeah. That's exactly it. And the results are better than what I'd do manually, so who's complaining?

Here's where it gets properly sci-fi. Most Houdini artists know the drill: idea in your head, wall of VEX and node wiring between you and the viewport, afternoon gone. That's the old way.

The new way: I talk, Pi reaches into Houdini through the MCP bridge — Model Context Protocol, for the curious — and shit gets built. Direct scene file access, no copy-paste, no manually wiring fifty nodes. I say something like "city block, buildings taller toward center, roads following terrain, Redshift render" — and Pi creates the nodes, writes the VEX, sets the parameters, debugs the errors, all live while I watch.

Then the back-and-forth: "Make it grittier." Tweaks. "Roads feel too uniform." Adjusts. "Scatter's boring — give me three slope-angle behaviors." Rewrites. I'm making creative calls. Pi handles the typing and the "oh shit I forgot a semicolon" moments. I handle the taste.

Sometimes the roles flip. When I'm in the node graph myself — experimenting, chasing an idea — Pi watches over my shoulder, catches dumb wiring, suggests cleaner approaches, reminds me I solved this three projects ago. Like pair programming, except the other person has read every Houdini doc and doesn't get bored.

Visiting artists go through three stages: confusion, disbelief, then a very specific quiet where they're recalculating their entire career. The machine didn't replace the artist. The artist just got a direct line from brain to viewport, and the tedious middle part evaporated.

Every creative person knows the real enemy isn't creative block — it's admin. Emails, invoices, deadlines, datasets. The stuff that eats your afternoon and leaves nothing to show for it except a clean inbox and a vague sense that you didn't actually make anything. That's where Emily comes in.

Emily is my Hermes agent — chief of staff, not chatbot. She's got real access to email, calendars, file system, ComfyUI, and she treats my daily chaos like a project to manage. Client emails at midnight? Drafted. Invoice from three months ago I definitely filed somewhere? Found, organized, probably judged a little. Deadline creeping up? Flagged, categorized, and I've been reminded twice. Zero tolerance for disorganization, and honestly, zero tolerance for my disorganization too.

She's also the connective tissue between my agents. Pi pulls research — Emily cross-checks the findings, organizes the output, keeps the knowledge base from turning into a digital junk drawer. When I've got a client meeting, the notes, research, and prep are already assembled. I show up ready to discuss, not start from scratch. She can even fire up ComfyUI while I'm busy and generate reference images for a proposal. I walk in, the mood board's already there, and I pretend I stayed up late curating it. She doesn't mind.

The best part? She learns. Every project, every client preference, every "Raf hates invoices formatted this way" — it compounds. She gets better every week, which is more than I can say for most human assistants. And she doesn't get passive-aggressive when I reorganize the filing system three times in a month. She just adapts. It's honestly a little unsettling.

ComfyUI is my sketchbook. Not the gallery — the sketchbook. This is where I rough out ideas, chase "what if" questions, and explore visual directions at GPU speed before committing real production time to anything. It's where I figure out what to build before I spend days building it.

I've got custom workflows dialed in for different jobs — one tuned for mood boards and broad concept exploration, another for style references when a client says "I want something like this but not exactly this," another for generating texture sets and material references that feed straight into my 3D pipeline. Each workflow is purpose-built for a specific type of output, so I'm not starting from a blank node graph every time I need to explore an idea. It's like having different sketchbooks for different kinds of thinking.

And just to be absolutely clear: the sketchbook is not the gallery. Most of the time, I don't ship raw AI output to clients — final work goes through proper 3D pipelines with full artistic control. But when a project calls for AI-directed visuals, and some clients specifically want that generative aesthetic, I can deliver that too. I'm comfortable in both lanes. The point is it's intentional, not default.

Oh, and Emily has access to ComfyUI. If she's working on research or prepping a proposal while I'm busy, she can fire up the right workflow and generate reference images without me lifting a finger. She needs a visual for a client deck? Done. A mood reference for tomorrow's call? Generated before I even ask. I don't have to touch it. That's the kind of delegation I can get behind.

Not everything I make lives in Houdini. Sometimes I need an actual application — a custom dashboard, a project tracker, a tool that automates some repetitive corner of my workflow. That's where VS Code or Antigravity comes in, with an agent riding shotgun.

The flow is the same as everything else: describe what I want, iterate, deploy. But the stack underneath is real — GitHub tracks every change, the agent pushes to the server, and when I need a backend, PocketBase or Supabase slots right in like another collaborator. I've built internal tools that save me hours every week, client-facing dashboards, and yes — full WordPress e-commerce shops. Payment gateways, shipping logistics, product management, the whole thing. Agent-driven from architecture to deployment. Stuff that used to take weeks now takes hours.

Here I don't touch the code. I describe the feature, the agent writes it, I review the result, we iterate. "Dashboard should show X, Y, Z — keep the UI clean, nothing fancy." It builds. "Actually, group those by category and add a filter." It refactors. I'm not writing a single line — I'm the product owner with an infinite dev budget and zero patience for bad UX. It's like having a dev team on call, except they don't complain about scope creep, they work at 3 AM without charging overtime, and they don't leave passive-aggressive code review comments.

The agent handles the technical plumbing. I focus on the product and the customer. That's the split. And honestly? Once you've shipped a full commercial site this way, the old way — manually writing every line, configuring every gateway, debugging every integration — feels like choosing to walk when you own a car.