Relevance AI sells the idea of an AI workforce, meaning agents that handle business processes with tools, memory and the ability to work together. The learning material is aimed at operations and business users rather than developers, and on accessibility it succeeds. Someone in a sales or support function can build a working agent from this documentation without writing code, which is a real achievement. The gap between that and the workforce framing is where the honest assessment lives.
The accessibility is genuine and worth crediting. Documentation written for people who do not think in code, with concepts explained in business terms and a builder that does not require technical background. Anyone who has watched a non technical colleague bounce off developer documentation will recognise how much work goes into this and how rarely it is done well. Tool and integration documentation is clear.
Connecting an agent to a data source, a CRM or an external service is explained step by step, and since most useful agents need to touch systems rather than just generate text, this material does real work. The templates are similarly useful for calibrating expectations, showing the size and shape of task these agents can reasonably handle. The multi agent framing is where I would push back. The idea of agents delegating to each other and coordinating on complex work sounds compelling and in practice coordination between agents compounds the reliability problems of each one.
An agent that is right ninety percent of the time, handing to another agent that is right ninety percent of the time, produces a chain with meaningfully worse odds, and the failure is often silent because each step produced something plausible. The documentation presents coordination as a capability rather than a risk, and anyone designing a multi step agent process needs to think hard about where the human checkpoints go. Error handling guidance is the biggest omission. What happens when an agent produces confidently wrong output, how you detect it, and how you build a process that catches mistakes before they reach a customer are the questions that determine whether this is useful or dangerous in a business setting.
The material covers building agents thoroughly and reviewing their work barely at all. That imbalance is common across this category and it matters more here because the audience is less equipped to spot the problem. Credit consumption is opaque until experience teaches you. Different operations consume different amounts, agent chains multiply the cost, and the relationship between what you build and what you pay is not obvious from the documentation.
Starting small and watching the consumption is the practical approach, and the material does not steer you towards it. Reliability expectations are set too high throughout. Agents work well on well defined repetitive tasks with clear inputs and predictable structure. They work poorly on ambiguous judgement calls, unusual cases and anything requiring real context about your business.
The documentation blurs that boundary, and a user who deploys against the wrong kind of task will conclude the technology does not work when the real error was in the selection. Three point four. Accessible, well written material that genuinely enables non developers to build useful automation, undermined by a multi agent narrative that promises more than it delivers and by almost no attention to what happens when the output is wrong.