AI agents are quickly moving from impressive demos into tools that may soon handle real business work. They can research, plan, write code, use software, make decisions, and complete multi-step tasks. But there is still a hard question sitting behind all that progress. How do teams know an AI agent is ready for the real world before it touches real customers, real workflows, or real money?
That is where the work of Miguel Monares and Jigsaw becomes interesting. Miguel Monares is building Jigsaw around a problem that could become one of the most important parts of the AI stack: creating simulated worlds where AI agents can train, fail, improve, and be evaluated before they are trusted with live work.
Instead of treating AI progress as only a matter of bigger models or better prompts, Jigsaw is focused on the environments around those models. In simple terms, the company is working on the training grounds that help AI agents learn how to act, not just how to answer.
Who is Miguel Monares
Miguel Monares is the CEO and co-founder of Jigsaw, an applied AI research company working on simulated environments, reinforcement learning environments, world generation, expert authoring, and agent evaluation. His work sits in a fast-growing part of artificial intelligence where models are expected to do more than generate text. They need to interact with tools, follow goals, handle uncertainty, and complete tasks in realistic settings.
Miguel’s background gives his founder story a strong research angle. He has been connected with work around AI agents, benchmarking, evaluations, and model behavior. That matters because the problem Jigsaw is addressing is not only a product problem. It is also a deep research problem.
Many founders are building applications on top of existing AI models. Miguel Monares is working closer to the layer that helps those models become more useful and reliable. That makes Jigsaw less like a simple AI tool and more like infrastructure for the next stage of agentic AI.
What Jigsaw is building
Jigsaw is building infrastructure for simulated environments where AI agents can be trained and evaluated. These environments can give agents a place to practice tasks, respond to changing situations, and prove whether they can perform reliably.
A static benchmark can tell you how a model responds to a set of questions. A simulated environment can test something deeper. It can show how an agent behaves over time, how it makes decisions, how it recovers from errors, and whether it can finish a task when the path is not perfectly predictable.
Jigsaw’s work is built around several connected ideas:
- RL environments where agents can learn through repeated interaction and feedback
- World generation that can create realistic settings for agents to operate in
- Expert authoring that allows people with domain knowledge to help shape tasks and scenarios
- Evaluation infrastructure that can measure how agents perform in dynamic situations
- High-fidelity environments that can better reflect real-world workflows
This is important because the next wave of AI agents will need more than good language skills. They will need judgment, planning, tool use, memory, error recovery, and the ability to adapt when conditions change.
Why simulated worlds matter for AI agents
For an AI agent, a simulated world can work like a flight simulator for a pilot. A pilot does not learn only by reading manuals. They practice in controlled environments where mistakes are safe, rare scenarios can be repeated, and performance can be measured.
AI agents need the same kind of training ground.
In the real world, an agent might need to process a procurement request, respond to a customer, write and test code, analyze a contract, manage a workflow, or coordinate across multiple tools. These tasks are not always simple. They involve context, tradeoffs, permissions, edge cases, and messy human instructions.
If an AI agent makes a mistake in a live environment, the cost can be serious. It might send the wrong message, approve the wrong action, miss a critical step, or create confusion inside a business process. Simulated environments help reduce that risk by giving teams a way to test agents before deployment.
They also make learning more scalable. Instead of waiting for real-world failures, teams can create thousands of training scenarios and let agents practice across them. The more realistic the environment, the more useful the feedback becomes.
The hidden bottleneck Miguel Monares is addressing
The AI industry often talks about model size, compute, and data. Those things matter, but they are not the whole story. As AI agents become more action-oriented, one of the biggest bottlenecks is the lack of rich environments where those agents can train and be evaluated.
A chatbot can be tested with question-and-answer prompts. An agent needs something more demanding. It needs tasks with goals, tools, constraints, changing information, and consequences.
That is the gap Miguel Monares appears to be targeting with Jigsaw. The question is not just whether an AI model can produce a smart response. The question is whether an agent can complete a job reliably across many situations.
This is a major shift. It moves evaluation away from simple output checking and toward behavior testing. In a simulated world, an agent can be judged on how it acts, not only on what it says.
That could make a big difference for frontier AI labs, enterprises, and developers who need more confidence before putting agents into production.
How Jigsaw connects research with real-world AI use
The most useful AI companies often sit at the intersection of research and practical need. Jigsaw fits that pattern. Its work is technical, but the reason it matters is easy to understand. Businesses want AI agents that can do real work, and they need better ways to train and evaluate them.
A company might want an AI agent to help with financial operations, customer support, compliance checks, software engineering, sales workflows, or internal research. Each of these areas has its own rules and expectations. A general-purpose model may understand the words, but that does not mean it understands the workflow well enough to act safely.
This is where simulation becomes useful. A realistic environment can reflect the steps, pressure points, and edge cases inside a workflow. It can help an AI agent learn what success looks like in context.
Jigsaw’s focus on expert authoring is especially important here. Domain experts know the details that generic training data often misses. They understand what can go wrong, what needs approval, what counts as a good outcome, and where human judgment matters. If those experts can help shape simulated tasks, the resulting training environments can become far more valuable.
Miguel Monares and the future of AI evaluation
AI evaluation is becoming one of the most important areas in artificial intelligence. In the early days of large language models, many evaluations focused on whether a model could answer benchmark questions. That still has value, but agents need a different kind of measurement.
An AI agent should be evaluated on whether it can complete tasks, use tools properly, plan across steps, notice mistakes, and adjust when something changes. This is much harder to measure with a static test.
Jigsaw’s approach points toward a more realistic future for agent evaluation. In a simulated environment, teams can test things like:
- Whether the agent completes the task successfully
- How often it gets stuck or takes the wrong path
- Whether it recovers after making an error
- How well it uses tools and external systems
- Whether it follows constraints and business rules
- How it handles uncertainty or missing information
- Whether it can perform consistently across many variations of the same task
This kind of evaluation could become essential as companies move from experimenting with AI agents to relying on them in daily operations.
Why Jigsaw’s approach stands out
Jigsaw stands out because it is not focused only on the surface layer of AI adoption. Many AI products are built around a chat box, a workflow shortcut, or a wrapper around a large model. Those tools can be useful, but they do not always solve the deeper problem of agent reliability.
Jigsaw is working on the infrastructure that helps agents become better before they are released into the real world. That is a more foundational bet.
The company brings together several areas that are becoming more important at the same time: reinforcement learning, simulation, world generation, evaluation, and domain expertise. Each piece matters on its own. Together, they create a stronger path for training agents that can handle real tasks.
This also gives Jigsaw a meaningful role in the broader AI ecosystem. If agents are going to become a major interface for work, then the environments used to train and test them will become a critical layer of the stack.
The role of domain expertise in simulated AI training
One of the biggest challenges in AI agent development is that real work is full of hidden knowledge. People inside a company often know things that are not written cleanly in a manual. They know which edge cases matter, which mistakes are costly, and which steps require careful judgment.
That kind of knowledge is difficult to capture with ordinary datasets.
Jigsaw’s focus on expert authoring suggests a future where domain experts can help turn their knowledge into training environments. Instead of asking an AI team to guess every possible scenario, experts can shape the tasks, rules, and outcomes that agents need to understand.
This could be useful in many fields:
- A procurement expert could help define realistic vendor negotiation scenarios
- A finance team could shape workflows around invoices, approvals, and reporting
- A software engineering team could create environments for debugging and code review
- A compliance team could build scenarios around policy checks and risk flags
- A customer support team could simulate difficult conversations and escalation paths
The value is not just in creating more data. It is in creating better practice. If the simulated environment reflects the real workflow, the agent has a better chance of learning behavior that actually transfers to business use.
The bigger shift toward agentic AI
The rise of agentic AI is changing what people expect from artificial intelligence. A model that only answers questions is useful, but an agent that can complete work has a different level of value.
That shift also raises the standard. If an agent is going to take action, it needs to be tested more carefully. A wrong answer is one thing. A wrong action inside a live workflow can create bigger problems.
This is why simulated environments may become a normal part of AI development. They give teams a way to test agents in controlled spaces, measure behavior, and improve performance before deployment.
Miguel Monares and Jigsaw are building directly into this shift. Their work reflects a belief that the next generation of AI will need stronger training grounds, not just stronger models.
What Miguel Monares’ journey says about the new AI founder playbook
Miguel Monares represents a newer kind of AI founder. He is not simply chasing a trend or adding AI features to an existing category. He is building around a serious technical bottleneck that becomes more important as the industry moves forward.
That kind of founder playbook is different. It requires research awareness, technical depth, and a clear sense of where the market is heading. It also requires patience, because infrastructure companies often do their most important work behind the scenes.
Jigsaw’s work may not always be as visible as consumer-facing AI products, but infrastructure often decides what becomes possible. Better cloud platforms helped software companies scale. Better developer tools helped teams build faster. In the same way, better simulated environments could help AI agents become more capable, safer, and more useful.
This is what makes Miguel’s story compelling. He is building in a part of AI that could quietly shape how agents are trained, tested, and trusted.
Why Jigsaw could become important as AI agents mature
As AI agents become more common, the demand for better evaluation will grow. Companies will not want to rely only on demos or benchmark scores. They will want proof that agents can handle real workflows under realistic conditions.
That is where Jigsaw could become important.
If Jigsaw can make it easier to create high-quality simulated environments, it could help AI labs and enterprises answer a practical question: is this agent ready to work?
The answer will not come from one test. It will come from repeated practice, better scenarios, stronger evaluation, and continuous improvement. That is exactly the kind of system simulated environments can support.
For enterprises, this could mean safer AI adoption. For frontier labs, it could mean better ways to improve and measure advanced agents. For domain experts, it could mean a more direct role in shaping how AI systems learn.
The next chapter for Miguel Monares and Jigsaw
Miguel Monares is building Jigsaw around one of the most important questions in modern AI: how can agents be trained and evaluated before they are trusted with real-world work?
By focusing on simulated worlds, RL environments, world generation, expert authoring, and agent evaluation, Jigsaw is positioning itself in a space that could become essential as AI agents move from prototypes into production.
The company’s work shows that the future of AI may not only depend on larger models. It may also depend on better environments. Agents need places to practice. Teams need ways to measure progress. Businesses need confidence before handing over important workflows.
That is the opportunity Miguel Monares is building toward with Jigsaw. He is not just building another AI product. He is helping shape the training ground where the next generation of AI agents could learn how to work.
