Project LASSO: Teaching AI to think ahead

Artificial intelligence has become remarkably good at answering questions. But there’s a difference between knowing what to say and knowing what to do.

Give an AI assistant a question, and it can produce an impressive answer. Give it a complicated objective that requires multiple decisions, changing circumstances, and an understanding of consequences, and things become much less reliable.

That’s the problem we’re exploring at COORDINATOR.

Our research project, LASSO (Latent Adaptable Strategic System One), investigates a different approach to AI decision-making: teaching systems to anticipate what might happen before deciding what to do.

Rather than relying exclusively on increasingly large language models, we’re exploring whether smaller, specialized neural networks can learn to understand their environments, predict consequences, remember experiences, and make better decisions over time.

Our latest experiments have produced encouraging results, exposed important limitations, and given us a clearer understanding of what it will take to build genuinely predictive AI.

The Problem: AI Can Act Without Understanding What Happens Next

Imagine asking an AI to accomplish a simple task:

Get to the other side of a locked door.

A human immediately understands that this might involve several steps. Find the key, pick it up, unlock the door, and walk through.

The important part isn’t knowing what a key looks like. It’s understanding why picking up the key matters.

The key has value because of something it makes possible later.

That’s a surprisingly difficult concept for AI systems to learn through experience.

Many AI systems can learn individual actions. Others can follow carefully written instructions or use predefined workflows. But learning how different actions contribute to a future objective—especially in unfamiliar situations—is a much harder challenge.

We believe that capability is essential for AI systems that can operate reliably beyond narrowly defined tasks.

Our Approach: Learn the World, Predict the Future, Choose the Action

LASSO combines several ideas from modern AI research into a single experimental system.

At its center is a learned world model: a compact internal representation of an environment and how it changes.

Instead of attempting to memorize every possible situation, the system learns patterns from experience.

It can then use those patterns to generate possible future outcomes.

A separate decision-making system evaluates those possibilities and helps determine which action to take.

Memory provides additional context from previous experiences.

The long-term objective is a system that can:

• Understand its environment through learned representations rather than hand-programmed rules.
• Anticipate consequences before committing to an action.
• Recognize dependencies between actions and longer-term goals.
• Learn from experience without continually starting from scratch.
• Transfer useful knowledge to unfamiliar situations.

We’re particularly interested in whether these capabilities can emerge from relatively small, efficient models.

The goal isn’t to replace language models. It’s to investigate whether specialized predictive intelligence can make AI systems more capable without requiring a larger language model for every decision.

That distinction could have significant implications for the cost, speed, and reliability of autonomous AI.

What We’ve Demonstrated So Far

Our research uses controlled environments where AI systems encounter unfamiliar layouts, obstacles, objects, doors, and tasks.

These environments may look like simple games, but they’re useful because they allow us to measure exactly what an agent has learned.

More importantly, they expose whether apparent intelligence comes from understanding the task or simply exploiting patterns in the training environment.

A Meaningful Improvement in Multi-Step Tasks

One of our latest experiments investigated whether LASSO could learn to complete tasks requiring several dependent actions.

The challenge was straightforward: navigate an environment, acquire a key, unlock a door, and reach the destination.

Our previous model completed none of the 64 locked-route challenges in either evaluation group.

We then compared two training approaches using the same underlying system and matched experimental budgets.

The results:

The consequence-focused approach substantially outperformed standard training on these tasks.

This is an encouraging result because the system wasn’t simply being evaluated on whether it could move through an environment. It had to complete a sequence of actions where earlier decisions enabled later ones.

However, these were development experiments, not independent confirmation of generalization. The improved model also failed our predefined advancement requirements, including one requirement to preserve performance on simpler tasks.

We are retaining those failures as part of the research record rather than presenting the model as a solved system.

Finding Where the Intelligence Breaks Down

Perhaps the most interesting discovery wasn’t the improvement itself.

It was what happened when the AI failed.

By analyzing complete decision sequences, we discovered that the improved model frequently managed to locate and approach the key.

But it often failed to pick it up.

Across the fresh set of 64 challenges:

47 times, the agent reached a position facing the key.

24 times, it successfully acquired the key.

21 times, it completed the entire objective.

And in every case where it successfully unlocked the door, it ultimately reached the destination.

This tells us something important.

For these particular tasks, the primary remaining obstacle isn’t navigating after the door is unlocked. It’s reliably recognizing and taking the intermediate action that makes success possible.

In other words, the system is beginning to perform multi-step tasks, but it has not reliably learned why certain actions matter to a future goal.

That gives us a much more precise research question:

How do we teach an AI to value an action because of what it enables later?

We believe answering that question could be an important step toward more capable autonomous systems.

Learning From the Past

Another part of LASSO investigates memory.

Today’s AI agents can store conversation histories, retrieve documents, and reference previous interactions. But storing information isn’t necessarily the same as learning how to use past experience to make better decisions.

We’re experimenting with systems that retain information about previous actions, predictions, outcomes, and mistakes.

The objective is to help future decisions benefit from what the system has already experienced.

In one controlled experiment, our agent successfully completed a challenging door-navigation task in 46 steps.

When we removed its accumulated recurrent history, the same model failed to complete the task within 320 steps.

That result demonstrates that historical context can materially affect the agent’s behavior.

It does not yet prove generalized memory-based reasoning, but it gives us a concrete example of why preserving and using experience matters.

Our longer-term research is focused on making these capabilities cumulative: allowing an agent to acquire useful knowledge, retain it, and build on it as new challenges arise.

An Important Discovery: Prediction Is Not the Same as Understanding

Our latest research has also exposed a limitation that we’re actively investigating.

LASSO can generate predictions of future states. But generating predictions doesn’t automatically mean those predictions are useful.

In some experiments, the learned forecasts were less accurate than a simple baseline that assumed the environment would remain unchanged.

Prediction errors also grew as the system attempted to look farther into the future.

Perhaps more importantly, we found that an agent can become dependent on the information produced by its predictive components without necessarily benefiting from physically accurate predictions.

That’s a critical distinction.

A system that performs well because it has learned to exploit the quirks of its own internal model isn’t necessarily a system that understands the consequences of its actions.

Our research is designed to separate those two possibilities.

We aren’t satisfied with an AI that merely appears to make intelligent decisions in familiar circumstances.

We want to establish whether it can use genuinely predictive knowledge to make better decisions when circumstances change.

That remains an open research challenge.

Why This Matters Beyond Games

The environments we’re using today are deliberately simple.

The underlying problem is not.

Consider an enterprise AI agent responsible for coordinating a complex workflow.

It might need to evaluate several possible actions, understand dependencies between tasks, anticipate downstream problems, and adjust its approach when new information becomes available.

A system that simply follows instructions can struggle when circumstances depart from the expected sequence.

A system capable of modeling consequences could potentially recognize those changes earlier and choose more appropriate actions.

The same principle applies to resource allocation, operations planning, scheduling, risk management, and coordination between multiple autonomous agents.

These are areas where knowing what action is available is often less important than understanding what that action will cause.

For COORDINATOR, this research has a particularly natural connection to our broader work.

Our enterprise platform focuses on governing AI agents: helping organizations define what agents are allowed to do, when they require human approval, and how their actions are evaluated.

LASSO explores a complementary question:

Can an AI system become better at deciding what it should do in the first place?

One addresses the boundaries within which agents operate. The other investigates the intelligence behind their decisions.

Together, they point toward autonomous systems that are not only more capable, but also more accountable.

A Different Direction for AI Research

Much of the AI industry’s progress has come from increasing model size, expanding training datasets, and investing in greater computational capacity.

Those approaches have produced extraordinary capabilities.

But we believe there’s another important direction worth pursuing: making AI systems better at learning how the world works and using that knowledge efficiently.

LASSO draws inspiration from predictive representation learning, world models, reinforcement learning, and memory-based decision systems.

None of those concepts is new on its own.

What we’re investigating is how to combine them into a practical, cumulative architecture that improves strategic decision-making.

We are particularly interested in architectures where specialized components handle prediction, memory, and decision-making without requiring a frontier-scale language model to perform every operation.

If successful, that could offer a different path toward increasingly capable AI—one that places greater emphasis on learned structure and predictive reasoning rather than computational scale alone.

It’s an ambitious hypothesis.

And we’re testing it experimentally.

What’s Next?

Our immediate priority is improving how LASSO connects predicted consequences to the value of individual actions.

The key-and-door experiments have given us a useful starting point.

We now want to determine whether the system can reliably learn why an intermediate action matters, apply that understanding across unfamiliar situations, and preserve its existing abilities while acquiring new ones.

Alongside that work, we’re investigating how to improve forecast accuracy, reduce prediction errors over longer horizons, and make historical experience more useful during decision-making.

Each experiment brings us closer to answering the central question behind LASSO:

Can we build an AI system that learns to think ahead, rather than simply learning what action to take next?

We don’t have a definitive answer yet.

But we’re developing the architecture, experimental evidence, and engineering foundation needed to find out.

And importantly, we’re learning not just what works, but why it works—and where it doesn’t.

Building the Next Generation of Autonomous Systems

COORDINATOR is building infrastructure for a future in which autonomous AI agents play an increasingly important role in business operations.

Our enterprise technology addresses the immediate need for governance, oversight, and control.

Our research investigates what those agents could become as their ability to understand, predict, and coordinate improves.

We see these as complementary parts of the same opportunity.

The near-term challenge is making AI agents safe and manageable enough for organizations to trust.

The longer-term opportunity is making them intelligent enough to handle increasingly complex decisions reliably and efficiently.

LASSO is one part of that journey.

We’re building toward AI that doesn’t just follow instructions. It learns from experience, anticipates consequences, and makes better decisions because it understands what comes next.

COORDINATOR Research — October 2026

LASSO is an active experimental research project. Results described here reflect controlled development experiments and should not be interpreted as production-ready capabilities or demonstrated general-purpose intelligence.