When Eric Ries popularized the concept of Minimum Viable Product, he established a potent benchmark for building software products people actually want.
The core insight of The Lean Startup was brilliantly simple – instead of assuming we know what users want, deploy the smallest thing possible and start learning from real user behavior.
The feedback-inducing MVP shaped the product management zeitgeist for a generation: When you're launching a social platform or productivity tool, for example, your biggest risk isn't technical feasibility – it's market fit.
Fast forward to 2025, and we find that launching AI agents presents a fundamentally different challenge, which exposes a key limitation in the MVP model and demands an additional validation measure.
When you're building an agent, your first hurdle isn't market fit – it's proving that your AI can perform assigned tasks reliably and appropriately. There's little point in testing whether users want your agent if you can't prove it does the job assigned to it.
Move over MVP and make room for the MVA, or Minimum Viable Autonomy. Before agents can walk to product market fit, they need to crawl to agent task competence.
Rethinking Risk in the Age of AI Agents
The MVP became axiomatic for product leaders by focusing on one crucial truth: most products fail not because they can't be built, but because no one wants them.
This insight helped transform software development from a lengthy waterfall process into rapid cycles of build-measure-learn. Companies could validate ideas in weeks instead of months, and "failing fast" became a virtue.
Reis asks us to consider Dropbox. Instead of building a complete file-synchronization system, they created a simple video demonstrating the intended user experience. The surge of interest from that video validated the core premise before they wrote much code.
Or we can look at Buffer, another popular lean startup case study, which began with a simple landing page to test if people wanted a tool for scheduling social media posts.
These MVPs weren't about testing technical feasibility – they were about validating demand. But AI agents present a fundamentally different validation challenge.
Imagine you're building a legal contract agent. You could create a beautiful interface and get potential users excited about the concept, but that wouldn't address your primary risks, which might be broken down as follows:
Can the system properly interpret legal language?
Can it identify potentially problematic clauses?
Can it provide useful explanations that make sense to lawyers?
These are questions of functional competence, not market fit.
Putting an agent in front of users before establishing competence is counterproductive. As Reis took pains to explain, "minimal viable" doesn't mean trivial, it means smallest useful, and it's the same when we're talking about autonomy.
What Minimum Viable Autonomy Looks Like
The key to a successful MVA lies in choosing the right scope. You want a domain narrow enough to be manageable but complex enough to demonstrate functional competence.
To understand MVA better, it helps to contrast it with more familiar approaches. You might initially think of this as a proof of concept (PoC) phase where your goal is to demonstrate basic agentic capabilities.
An MVA in our way of thinking, however, goes further than a typical PoC. The goal here is not just proving that a technology can work; it's validating that an AI system operates effectively in a controlled but realistic environment.
It's not enough to show that an agent executes scripted operations – an MVA validates the quality and appropriateness of those operations and the system's ability to provide useful explanations for its outputs.
Where PoCs often rely on idealized or simulated environments, an MVA operates under controlled but more realistic conditions, requiring the system to handle variations within its domain rather than following only predetermined paths.
Think of it this way: a legal assistant PoC might show that an AI can identify contract clauses in carefully selected examples. An MVA demonstrates that the AI can identify problematic clauses across various contracts, provide useful explanations, improve through feedback and indicate when it needs human expertise.
The MVA represents a higher bar for validation – not just "Can the technology work?" but "Can this system perform its assigned tasks competently?"
Building Your First MVA
To get our heads around building an MVA, let’s consider three different approaches to building an AI customer service assistance system.
A PoC approach might create a demo showing the AI responding correctly to five common customer inquiries. This proves the underlying technology works but says nothing about performance under varied conditions.
An MVP approach might create a chatbot interface with basic predefined responses and gather user feedback. This would tell you if customers want AI support, but not whether the AI can actually provide effective assistance.
An MVA approach instead focuses on proving functional competence in one specific type of interaction – perhaps handling refund requests for a single product category.
This scope allows you to validate the system's ability to understand customer queries, apply guidelines, suggest actions, improve with feedback and recognize situations beyond its capabilities – all in controlled tests that simulate real interactions.
Three Patterns for Measureing Success
When measuring the success of an agent, we need to move beyond traditional software metrics like user adoption and engagement rates. These don't capture what really matters here: the quality of performance and ability to improve.
Instead, we need to evaluate three core patterns that define agentic potential.
Decision Quality: Beyond Simple Accuracy
Raw accuracy can be misleading in AI systems. A system that's 99% accurate but provides no explanation for its outputs or indication when it's outside its training might be worse than one with 90% accuracy that communicates confidence levels and reasoning.
Consider our legal document review system. Beyond identifying potentially problematic clauses, we need to evaluate several dimensions of performance:
Does it flag clauses for appropriate reasons?
Can it provide explanations that legal professionals find useful?
Does it recognize situations outside its experience?
Is its performance consistent across similar cases?
The goal isn't flawless accuracy – it's reliable, explainable and appropriate operation.
Learning Rate: The Heart of Autonomy
A crucial difference between automated systems and AI agents is the ability to improve over time through human feedback, structured training and outcome analysis.
In our legal system example, learning happens when legal experts provide feedback on recommendations, when the system processes more examples and when it incorporates data about which of its suggestions were accepted or rejected.
We're measuring how efficiently the system incorporates these signals. Along these lines, some key questions could include:
How much feedback does the system need to handle new contract types?
Can it apply patterns from one domain to another when given appropriate examples?
How much guidance is needed before it stops flagging similar non-issues?
Think of it like training a legal associate. Their learning rate isn't about independent improvement but how efficiently they incorporate guidance to boost performance.
Similarly, an AI system's learning rate measures how well it translates various inputs into improved capabilities.
Operational Reliability: Performance Boundaries
The third critical pattern is how well the system performs within its defined scope.
This also isn't about raw performance metrics – it's about appropriate, reliable and predictable operation. Looking again at our legal assistant example, key indicators include:
Does is maintain consistent performance across different inputs?
Can it handle reasonable variations appropriately?
Does it recognize when a case falls outside its capabilities?
Can it provide useful information about its decision process?
Operational reliability also includes the system's ability to manage challenging cases. While current AI systems may not be able to truly "prioritize" tasks, they can be designed to handle varying workloads, triage requests based on importance and maintain performance quality even when processing complex inputs.
Think of operational reliability like evaluating a team member. You're assessing not just their output quality, but their judgment about when to ask for help and their ability to work within established parameters.
Bringing It All Together
The art of measuring MVA success lies in balancing these three patterns.
A system might score high on decision quality but show poor learning rates. Another might learn quickly but demonstrate unreliable performance. The goal is finding the right balance that proves genuine functional capability within your chosen scope.
This is why MVA metrics need to be both quantitative and qualitative. Numbers matter, but so do patterns, trends and the quality of outputs when evaluated by domain experts.
Success in an MVA isn't just about hitting metrics – it's about demonstrating reliable, improvable capability in your chosen domain.
The Path Forward
The MVA framework represents an evolution in how we build AI systems. It acknowledges that before we can validate market fit, we must first validate functional competence.
This two-stage approach – first proving capability, then testing market fit – recognizes the truly special nature of agentic software and provides a clearer path to developing agents that succeed with toughest test of all: actual users.
As we continue to develop more sophisticated AI systems, the core insight will remain: the first risk in building AI agents isn't that no one will want them, but that they won't be capable of performing as expected.
By focusing first on proving capability in a narrow but meaningful scope, we can build AI systems that don't just promise assistance, but actually deliver it.
Agentic Foundry: AI For Real-World Results
Learn how agentic AI boosts productivity, speeds decisions and drives growth
— while always keeping you in the loop.