$75M founder reveals his Agentic Engineering setup

Summary

The discussion centers on how 10X (co-founded by Alex Lieberman) approaches “agentic engineering”, and how AI reshapes both software development and enterprise consulting. The core idea: in an AI-native world, software work will rely less on developers manually reading/writing code, and more on systems that provide agents with structured context, guidance, and validation.

1) AI-native transformation for old, slow companies

To become AI native, 10X argues companies must meet organizations where they are in their AI journey and integrate AI across the entire value chain—not just deliver a solution and leave.

They contrast their approach with traditional consulting, where older firms may deliver long slide decks (e.g., 150 pages) and then walk away, leaving clients to implement.

10X’s model is a “full-stack partner”, including:

• Teaching leadership to use tools (e.g., Claude/Copilot-like tooling referenced as “Claude code” and “Clawed co-work”)
• Implementing agents for specific business functions (examples: invoicing, AR/AP, HR back-office workflows)
• Driving change-management and ROI, not just strategy

2) “Innovation as a service” + implementation + compounding leverage

They argue that genuine consulting differentiates by providing not only ideas, but implementation and measurable business outcomes.

Their framing:

• Enterprise-grade, applied AI innovation
• Engineering teams “forward deployed” into messy, high-stakes problems
• Diversity of work as an advantage: engineers can work on multiple client problems rather than staying stuck on one narrow internal project long-term

3) Single-player vs multiplayer AI (where leverage grows)

They propose a client framework:

• Single-player AI: place strong models in front of individuals (e.g., “zero behavioral change” workflows like asking questions the same way people used Google). Useful, but limited impact because it doesn’t compound.
• Multiplayer AI: redesign workflows/processes that affect whole teams/functions (e.g., sales agents, horizontal operational automation).

They also note that clients often arrive with “AI problems” that are actually data readiness problems, requiring data engineering before agents can work effectively.

4) “Context as code” and the engineering setup for agents

A major theme: the biggest limiter for agent performance is not the model—it’s coordination, memory, and coherence across tasks and sessions.

Key claims:

• Traditional software treats the codebase as the primary artifact.
• 10X treats structured documentation/context as equally crucial—or more crucial.
• They use a “meta harness” (internal term) / context repo containing markdown-based artifacts agents can reliably read and follow.

Their SDLC is redesigned for agents:

• Agents require structured artifacts (e.g., epics, specs, docs, conventions)
• Every significant work session writes back to a log and updates project state (e.g., Linear tickets)
• A system provides the right information at the right moment using “benevolent prompt injection” at agent startup

5) Validators and drift detection (“linting” for the whole process)

They describe validation tooling that checks consistency across documentation/process artifacts—similar to linting, but applied to the full SDLC.

Example:

• A project spec “checkpoint” marked complete disagrees with a derived status computed from rules (some specs/tickets aren’t actually complete yet).

They suggest this enables a self-improving loop where agents detect drift and fix it next.

6) Their content machine: scaling human perspective into abundant output

They discuss a system for turning employees into consistent content creators while avoiding “AI slop.”

Thesis:

• In a post-AI world, business moats shift toward people/talent and trusted distribution (audience trust becomes revenue).

10X’s “content machine” aims to create content abundance without burning out humans.

System structure

• Process layer (in git): pipeline steps for generating, editing, scoring, repurposing, and distributing content
• Personal layer (on the creator’s computer): “voice” files and content lessons so output matches the individual’s perspective and style

Workflow highlights

• Creator select: the human chooses the idea and provides their own words (AI shouldn’t replace the origin of thought)
• Research/synthesis steps, followed by a multi-person “interview panel” style questioning to force specificity (persona-like skills)
• A “writing/editor council” scores quality; if below threshold, it triggers a revision loop
• A repurposing engine turns one piece into many derivative posts, again through quality checks
• Distribution includes UTM tagging and performance feedback; feedback improves future output

They claim the antidote to “AI sameness” is that the origin is human, while AI mainly improves flow, editing, and completeness.

7) Debate-adjacent claim: “Will nobody read or write code?”

They discuss a hot take (attributed to Matt Van Horn) that soon nobody (including engineers) will read/write code.

Their response is nuanced:

• Code-reading may decline, but engineers still need deep fundamentals to understand and structure systems
• AI can amplify weak habits in weak engineers; better agent fluency and structure can outperform raw coding skill
• Agents should start each session with full context so humans aren’t the only “memory layer”

8) How their agent engineering setup works (walkthrough)

They separate repositories:

• Code repos
• A project management repo containing structured artifacts

They use convention artifacts (files with a “con” prefix) with metadata and indexes so agents can parse conventions.

They provide a CLI workflow that exposes project artifacts/status differently depending on mode:

• Operator mode (human-facing)
• Agent mode (more detailed artifact listings, including logs and where to find relevant context)

Each agent session begins with a startup hook that injects the correct packet of context and instructions, enabling the agent to act like a “senior engineer” immediately.

9) Closing: open sourcing and where to learn more

They don’t fully commit to open sourcing, but discuss an “open discussion” about what might be released later.

They point viewers to:

• 10X for careers
• Alex’s social account for ongoing content

Presenters / contributors

• Alex Lieberman (co-founder of Morning Brew and 10X; presenter)
• Dan (Director of Engineering at 10X; presenter)
• David Andre (podcast host; appears in the dialogue)
• Arman (Alex’s co-founder / co-managing partner of 10X; referenced)
• Matt Van Horn (referenced; interviewed previously)
• CJ (10X engineer; referenced)
• Peter Yang (referenced via an “anti-AI slop” skill example)
• Morgan Hel, Tim Urban, Sean Pur, Greg Eisenberg, David Prell (referenced as “editor council” persona skills)
• Barbara Walters, Larry King, Michael Barbaro, Joe Rogan, Tim Ferriss, Howard Stern (referenced as “interview panel” persona skills)