Building Agents • aisdk

Agents are AI entities with a role, a goal, and (optionally) tools. In aisdk, agents are built on top of the R6 object-oriented system.

Creating Simple Agents

You can create an agent using create_agent().

library(aisdk)

# Create a Math Agent
math_agent <- create_agent(
  name = "MathAgent",
  description = "Performs arithmetic calculations and mathematical reasoning",
  system_prompt = "You are a math assistant. return only numerical results."
)

print(math_agent)

Running an Agent

To run an agent, use the $run() method. You need to provide a model (or string ID).

model <- create_openai()$language_model("gpt-4o")

result <- math_agent$run(
  task = "What is the square root of 144?",
  model = model
)

print(result$text)

Shared Session State

Agents can share memories and environments through a ChatSession. This is powerful for multi-agent workflows.

# Create a shared session
session <- create_chat_session(
  model = model,
  system_prompt = "You are a helpful assistant."
)

# Store some shared data
session$set_memory("project_name", "Apollo")
env <- session$get_envir()
env$data <- data.frame(x = 1:10, y = rnorm(10))

# Run agent with session context
# The agent can now access 'data' if it has tools to do so, 
# or conceptually if we inject it into the prompt.
result <- math_agent$run(
  task = "Analyze the project data",
  session = session,
  model = model
)

Specialized Agents

aisdk comes with pre-built specialized agents:

create_coder_agent(): Optimized for reading and writing code.
create_planner_agent(): Optimized for breaking down tasks.
create_visualizer_agent(): Optimized for creating plots (ggplot2).

coder <- create_coder_agent()
visualizer <- create_visualizer_agent()