“Agent Harnesses Are Dead. Long Live Agent Harnesses.” — João Moura

Why this is in the vault

Dissent #6 in the harness thesis research. Moura is the founder of CrewAI — he builds multi-agent frameworks, so he has skin in the game. His critique of the “harness as moat” thesis is the strongest one we’ve seen yet because he proposes a specific alternative: entangled software. This also has direct implications for RDCO’s consulting positioning and the engagement model for helping clients build their own agentic systems.

The core thesis

Frameworks got commoditized. Now harnesses are getting commoditized. The vocabulary cycles — “frameworks → scaffolds → harnesses” — but the underlying dynamic hasn’t changed. Every quarter model providers absorb another primitive into the API. The debate about what to call the building tools is “furniture rearrangement.”

Moura’s line: “The harness should be thin. The harness is plumbing. Plumbing matters, but nobody builds an exciting defensible product on plumbing.”

Notably, he explicitly agrees with Garry Tan’s “thin harness” position — he’s not disputing Tan. He’s disputing the idea that thin harnesses are the new moat. Chase’s “your harness, your memory” thesis is his specific target.

What Moura says is the real moat

When building gets cheap, value moves to what can’t be replicated overnight:

Distribution
Proprietary data that took years to assemble
Products that capture intelligence as customers use them and get better over time
Trust earned through production

The killer line: “You can’t vibe-code the thousandth customer’s accumulated patterns feeding back into the product. That flywheel is earned, not built.”

Entangled Software (the new concept)

The novel contribution. Borrowed from quantum physics — when two particles become entangled, the state of one instantly reflects the state of the other.

Applied to software: the product and the customer influence each other. The customer’s behavior shapes the software. The software shapes how the customer works. Over time, they become inseparable.

“This is the opposite of how software has worked for thirty years. We’ve always built tools and asked humans to adapt to them. Entangled software flips that. The software adapts to behavior instead of behavior adapting to software.”

Key claim: this wasn’t possible before agents, but now is. “Entangled agentic systems” are the next phase — where agents don’t just execute tasks, they learn from every customer’s workflows, adapt to how each organization actually operates, and get better the more they’re used.

His quote: “We’re building toward a world where you don’t set up agents, they emerge from how your team works.”

The “road, not the car” metaphor

The framework/harness debate is about how to build the car. Moura says the companies that win won’t be the ones with the best car — they’ll be the ones that built the road. “The infrastructure of trust, data, and adaptation that every car needs to drive on.”

Mapping against Ray Data Co

Uncomfortable truth #1: Our current positioning (horizontal harness + skills for clients) is exactly what Moura says is NOT the moat. If we sell “here’s the framework to build your own AI COO,” we’re selling the car, not the road. That’s a commodity position over time.

Uncomfortable truth #2: RDCO internally IS entangled software for the founder. The vault learns from his work. The skills evolve from his feedback. The system gets better with use. This is demonstrably entangled. Nobody else has a system that knows the founder’s voice, business context, feedback patterns, and operating rhythm.

The real RDCO moat, per Moura’s framework:

NOT the skills (they can be copied)
NOT the vault structure (it can be cloned)
YES the specific entanglement — 1,415+ documents of founder-curated assessments with specific cross-references, decisions, feedback patterns, and domain expertise that compound only when THIS founder uses THIS system

Consulting implications:

When we help clients build their COO agents, the value is NOT the framework install
The value is the entanglement that grows over months/years of use with their specific data, processes, and patterns
Engagement model: “plant a seed and nurture it” — long-term, stewardship-based, not project-based
Pricing: retainer + evolution, not one-time build
Differentiation from MG-style consulting: we don’t deliver a deliverable and leave; we co-steward the entangled system

Content implication:

This frames the MAC content strategy differently. MAC isn’t just a lead magnet for consulting — it’s a framework that BECOMES entangled with each company that uses it. The test matrix fills in differently for each business. The severity calibration drifts with the team’s tolerance over time.
Sanity Check becomes the entry point to a long-term entangled relationship, not a subscription product.

Where Moura’s argument is weakest

The entanglement argument is hard to verify. How do you know when software is entangled vs just “customized”? CRMs and ERPs have been “adapting to customer behavior” for 20 years through configuration. What’s genuinely new here?
CrewAI positioning coloration. Moura is pitching CrewAI as the company doing entangled software. The article is simultaneously a strategic essay and a product roadmap tease. Some of the abstract framing may be retrofitting CrewAI’s recent pivot.
Doesn’t address the open-source commoditization path. If entanglement is the moat, does an open-source CrewAI clone + someone else’s data become the moat? He doesn’t address this.

synthesis-harness-thesis-dissent-2026-04-12 — the original 5 dissents; this becomes #6
2026-04-12-cross-check-agent-architecture — the cross-check report that should be updated
2026-04-11-garry-tan-thin-harness-fat-skills — Tan’s position that Moura agrees with
2026-04-11-harrison-chase-your-harness-your-memory — Chase’s “harness is the moat” that Moura specifically disputes
2026-03-31-semistructured-data-layer-does-the-work — Natkins’s “data is the moat” — closest to Moura but weaker framing
2026-04-11-jaya-gupta-anthropic-sees-moat — Anthropic’s moat thesis from the platform side