The model is a commodity. The infrastructure around it is the real work.
We've been building AI infrastructure in production for months: agent teams, knowledge systems, automated intelligence pipelines. This is where we share what we're finding. The architecture, the failures, and the patterns that actually hold up.
Mainline AI is a research and engineering hub for AI-first infrastructure. Building in public.
The money is moving. The results aren't.
Organizations doubled their AI spending in 2025. The mandate is coming from the CEO. But most companies are stuck, not because the tools aren't ready, but because individual AI usage doesn't become organizational capability on its own.
The gap is infrastructure. The coordination layer, the memory, the feedback loops, the process redesign. Almost nobody is working on this layer seriously. We are.
What we believe
AI organizations, not AI individuals
Individual productivity gains don't compound unless the organization is redesigned around them. The goal isn't faster employees. It's a faster organization.
The harness beats the model
The same model produces wildly different results depending on how it's deployed. The infrastructure around the model matters more than the model itself.
Working systems, not strategy decks
The measure of progress is something running in production, not a slide deck recommending future work.
Process engineering is the technology
The real work is redesigning how decisions get made, how information flows, and where judgment lives. The AI tooling follows from that.
Outcomes over activity
What matters is whether the organization actually changed. Not how many hours were spent, not how many tools were evaluated. Did the system get better?
Intelligence automates, judgment stays human
AI should own the repeatable, the routine, the high-volume. The strategic, the ethical, the relational: those stay with people. This is the line worth drawing carefully.
The market is arriving at the same conclusion.
Our research tracks the industry closely. In just the past quarter, VCs, enterprise CEOs, ML researchers, and frontier lab leaders have been converging on the same thesis: the infrastructure around the AI matters more than the AI itself.
“A large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge.”
Andrej Karpathy, April 2026
Former Director of AI at Tesla, OpenAI founding team
“There is likely a massive capability overhang in most enterprises. The number one thing needed is a new form of system integrators and consultancies.”
Aaron Levie, March 2026
CEO of Box
““Harness infrastructure” is Frontier #1. 78% of AI failures are invisible. Memory, context, and evaluation are the new differentiation layer.”
Bessemer Venture Partners, April 2026
"Five Frontiers of AI"
“The harness is what distinguishes production agents from demos. Feedforward guides, feedback sensors, computational and inferential controls.”
Martin Fowler, April 2026
"Harness Engineering for Coding Agent Users"
We've been building this infrastructure since February. Not because these people told us to. Because we saw the same problem from inside a 100-person engineering org and decided to do something about it.
Harness > Model
The model is the engine. The harness is the car, the road, and the GPS. Nobody buys an engine.
The same model scores 42% or 78% on the same benchmark depending solely on the infrastructure around it. The harness is everything: the context it receives, the tools it can use, the memory that persists between sessions, the feedback loops that catch drift, the scheduling that compounds intelligence while you sleep.
This is the layer that turns a chatbot into a production system. We've been building and studying this layer for months. Here's what we've found.
Kevin McNamee
Engineering Leader · Researcher · Builder
15 years building and leading engineering teams. Currently leads 100+ engineers at FanDuel, shipping real-time pricing systems serving millions. Previously VP of Engineering at Bowery Farming.
Built a production AI knowledge system for his own work: 8 specialist agents, 15 automated jobs running 24/7, 60+ analyzed intelligence pieces compiled into a persistent knowledge base with integrity checking, trend tracking, and overnight synthesis. Not a side project. The system he depends on daily.
Writes about what he learns. The findings, the failures, and the patterns that actually hold up when you run them in production for months, not minutes.
What this looks like in practice
These aren't mock agents. They run daily on a headless Mac Mini. Each has a defined role, identity file, and context layer tuned over months of production use.
Kevin McNamee
Leader
Human
Philly
Main Agent
Claude Opus
Lisa
Curator
Claude
Frink
Researcher
Claude
Edna
Reviewer
Claude
Ned
Coder
Codex
Martin
Autoresearcher
Claude
Codex
Implementation
OpenAI
Claude Code
Architecture
Claude
Test Runner
Validation
Vitest
The pattern repeats. Every executor can lead their own team.
What we built and run daily
A production AI knowledge system running on a headless Mac Mini since February 2026. This is our research lab and daily operating system.
60+
Intelligence pieces analyzed
18
Reusable pattern cards extracted
15
Automated jobs running 24/7
8
Specialist agent roles
40+
Days of continuous memory
9
Integrity checks every morning
The system ingests articles, podcasts, and research. It compiles them into structured knowledge with cross-references and quality gates. It tracks trends across 60+ sources and upgrades patterns to confirmed trends when evidence hits critical mass. It catches drift before it compounds.
Agents don't just execute tasks. They have identities, conventions, and quality standards. The curator writes differently than the researcher. The reviewer has different priorities than the coder. Specialization is functional, not cosmetic.
One piece of this system became a product.
Imprint is a composable agent identity system: a CLI and visual builder for creating the identity files that make agents specialists instead of generalists.
imprint.getmainline.aiWriting
Practice, not theory. Everything comes from building.
The Harness Is Where the Leverage Lives
March 2026
The model is a commodity. The infrastructure around it is the competitive advantage. A thesis developed through building, not just observing.
Read on XI’ve Been Running a Karpathy-Style Knowledge Base for a Month
April 2026
Field report from running the exact architecture Karpathy described, with 8 specialist agents and 15 automated jobs. What actually works.
Coming soon