The rise of AI agents with João Moura of CrewAI
Summary
CrewAI is betting that enterprise value will accrue to the control plane, not just the agent-building framework. Moura expects companies within three years to operate “thousands if not hundreds of thousands” of agents, creating demand for planning, building, deployment, monitoring, integration, authentication and access controls. Otherwise, those agents become thousands of disconnected legacy codebases rather than a manageable corporate asset.
Adoption begins with low-precision, human-reviewed work and can gradually expand toward more decision-making. Observed use spans sales, marketing, back-office processes, coding, IRS forms, media editing and runtime pricing. Moura’s formula for eventual value is task complexity multiplied by autonomy, but “we’re not in that top quartile just yet” where critical processes run without hand-holding.
Executive sponsorship, technical ownership and a clear pain point separate credible enterprise buyers from agent tourists. A company asking CrewAI what it should automate is usually a weak signal; successful customers arrive knowing the process they want to change and have engineers able to unlock internal systems. Some Fortune 500 customers progressed from simple workflows to agents adjusting marketplace prices while monitoring competitors.
Scale is accelerating, but CrewAI constrains complex work through specialization rather than unlimited autonomy. Moura said January alone was over 50 million agents and described a 21-agent crew producing competitive-intelligence reports tens of pages long, with separate specialists for finance, branding, positioning and marketing. For a roughly 60-page IRS form backed by a 720-page manual, however, code isolates one question at a time before agents research and answer it.
The agent economy pressures both software pricing and model economics. Per-seat SaaS becomes awkward when one highly productive agent can eliminate the need for a human seat, yet Moura expects agent-specific endpoints to arrive later than many assume because teaching agents existing human interfaces “unlock[s] the entire internet.” Meanwhile, o1 and o3 are capable but have not taken off much in agent workloads: builders prefer the cheapest, fastest model that clears the task, including GPT-4o mini.
Open source has strategic momentum, although closed models still carry most CrewAI workloads today. Regulated finance and insurance customers increasingly self-host models in an “airtight container,” while R1 has produced mixed US-adoption signals but strengthened Moura’s belief that open models can advance further. He predicts renewed fine-tuning and smaller models outperforming today’s 70-billion-parameter systems for agent workloads.
Near-term labor impact looks more like organizational leverage than full replacement. Moura sees teams moving from four people doing repetitive work to one supervising agents while three are reallocated, with more agent involvement in decision-making perhaps arriving over the next couple of years. His hedged advice is to learn the tools: that may or may not be enough, but statistically it should improve one’s position.
Deep dive
Not yet available upstream; scheduled sync will retry.