What We Learn From AI Agents - Part 1

AI Agents are an important part of the Victoria VR AI Hub because they allow us to generate and analyze gameplay data across many different matches, configurations, and strategies.

Each Agent can behave differently. It can use different spells, loadouts, positioning, combat logic, and levels of aggression or risk.

When many different Agents play the game, we can start identifying patterns that would be difficult to see from a small number of manually played matches.

The value is not simply that AI Agents play the game.

It is what we can learn from the way they play.

What AI Agents Help Us Understand

One of the most useful areas of AI Agent data is combat behaviour.

Different Agent configurations allow us to compare spells, builds, strategies, and reactions across a wide range of situations.

Spell Usage

We can see which spells are selected most often and which ones are rarely used.

If one spell appears in a very large percentage of Agent configurations, it may indicate that it is particularly useful or versatile.

At the same time, if another spell is consistently ignored, we can investigate why.

It may be too situational, difficult to use effectively, or simply less attractive than other available options.

The important signal is not one Agent making one decision. It is seeing the same behaviour appear repeatedly across many matches and configurations.

Spell Combinations

Individual spells are only one part of the picture.

AI Agent matches can also show us which combinations of spells work particularly well together.

Certain abilities may become much stronger when paired with another spell, while other combinations may provide very little advantage.

A spell that appears balanced on its own can behave very differently when combined with another mechanic.

By comparing these combinations across many matches, we can better understand how abilities interact with each other.

Loadout Performance

Different Agents can enter matches using different complete loadouts.

This allows us to compare entire builds rather than looking only at individual abilities.

If a particular loadout repeatedly performs well against different opponents and strategies, it may deserve closer attention.

The same applies to builds that consistently struggle.

Some loadouts may perform extremely well in specific situations but poorly against certain strategies. Others may remain effective across many different conditions.

This gives us a broader picture of how complete builds perform.

Comparing Different Strategies

Because Agents can behave differently, we can compare different approaches directly.

An aggressive Agent can play against a defensive one.

A close-range strategy can compete against a configuration designed to maintain distance.

A high-risk Agent can be tested against one that prioritizes survival and positioning.

This helps us understand not only which strategy wins, but also when and why it performs well.

A strategy may be strong against one type of opponent while struggling against another.

That information is often more useful than simply looking at the final result of a match.

Reactions to Combat Situations

AI Agent data can also show us how different configurations react when circumstances change.

We can observe things such as:

• What happens when an opponent gets close
• Whether an Agent retreats under pressure
• How it repositions after using an ability
• When it chooses to attack
• When it chooses to defend
• How it responds to different opponent strategies

This gives us another layer of information beyond wins and losses.

An Agent that performs well in one situation may react very differently once it is placed under pressure or forced out of its preferred position.

Dominant and Underused Mechanics

One of the most useful signals appears when many different Agents independently start relying on the same mechanic.

If different configurations repeatedly choose the same spell, combination, or strategy, that may indicate unusually high value.

The opposite can also be important.

If a mechanic is rarely used across many different configurations, we can investigate why it is being avoided.

Neither result automatically means something needs to be changed.

The data simply tells us where it may be worth looking more closely.

From Individual Matches to Patterns

A single Agent match gives us one example.

A large number of matches across many different configurations allows us to start identifying broader patterns.

We can ask:

• Which spells are consistently preferred?
• Which abilities are rarely used?
• Which combinations repeatedly perform well?
• Which loadouts struggle?
• Which strategies work against which opponents?
• Which mechanics appear unusually influential?
• How do different configurations react under pressure?

This is where AI Agents become especially useful.

They allow us to examine combat behaviour across a much wider range of situations and strategies.

In Part 2, we look beyond individual combat decisions and explore what AI Agent data can tell us about movement, maps, encounter locations, and pacing.

The value of AI Agents is not simply that they play the game. It is the patterns their gameplay can reveal.