State

Margarita supports storing and accessing state during a run.

@state count = 0

<< What is the count variable? >>

@effect run

The example above defines a state variable named count.

To read a value from state, instruct the agent to "use the count variable" or "get count from shared state"; the system will return the current value.

Note that language models can be nondeterministic, so the agent may not always call the state tool.

You can also instruct the agent to set or update state values. For example, asking it to "set count to 5" or "update count to 10" will update the count variable in state (again, the agent may not always call the tool due to model nondeterminism).

@state count = 0

<< Set the count variable to 5 >>

@effect run

The example above sets the count variable to 5.

Setting values in the context from state

State variables can also be filled into the context before a run. To do this, use ${var_name} syntax in your template.

@state name = "Margarita"

<< Hello, ${name}! >>

This will fill out the name variable from state and include it in the context before the run, so the agent doesn't have to use a tool call to set/get it.

This can be helpful for reducing the number of tool calls the agent has to make which saves you tokens. It also gives you more control since you can set the values into the context before a second run.