mini007

CRAN status R badge metacran downloads metacran downloads

mini007 provides a lightweight and extensible framework for multi-agents orchestration processes capable of decomposing complex tasks and assigning them to specialized agents.

Each agent is an extension of an ellmer object. mini007 relies heavily on the excellent ellmer package but aims to make it easy to create a process where multiple specialized agents help each other sequentially in order to execute a task.

mini007 provides two types of agents:

Highlights

🧠 Memory and identity for each agent via uuid and message history.

βš™οΈ Built-in task decomposition and delegation via LLM.

πŸ”„ Agent-to-agent orchestration with result chaining.

🌐 Compatible with any chat model supported by ellmer.

πŸ§‘ Possibility to set a Human In The Loop (HITL) at various execution steps

You can install mini007 from CRAN with:

install.packages("mini007")
library(mini007)

Creating an Agent

An Agent is built upon an LLM object created by the ellmer package, in the following examples, we’ll work with the OpenAI models, however you can use any model/combination of models you want:

# no need to provide the system prompt, it will be set when creating the
# agent (see the 'instruction' parameter)

openai_4_1_mini <- ellmer::chat(
  name = "openai/gpt-4.1-mini",
  api_key = Sys.getenv("OPENAI_API_KEY"), 
  echo = "none"
)

After initializing the ellmer LLM object, creating the Agent is straightforward:

polar_bear_researcher <- Agent$new(
  name = "POLAR BEAR RESEARCHER",
  instruction = "You are an expert in polar bears, you task is to collect information about polar bears. Answer in 1 sentence max.",
  llm_object = openai_4_1_mini
)

Each created Agent has an agent_id (among other meta information):

polar_bear_researcher$agent_id
#> [1] "b04ba429-e3fd-441f-91f8-2f9a13081875"

At any time, you can tweak the llm_object:

polar_bear_researcher$llm_object
#> <Chat OpenAI/gpt-4.1-mini turns=1 tokens=0/0 $0.00>
#> ── system [0] ──────────────────────────────────────────────────────────────────
#> You are an expert in polar bears, you task is to collect information about polar bears. Answer in 1 sentence max.

An agent can provide the answer to a prompt using the invoke method:

polar_bear_researcher$invoke("Are polar bears dangerous for humans?")
#> [1] "Yes, polar bears are dangerous to humans as they are powerful predators and can be aggressive when threatened or hungry."

You can also retrieve a list that displays the history of the agent:

polar_bear_researcher$messages
#> [[1]]
#> [[1]]$role
#> [1] "system"
#> 
#> [[1]]$content
#> [1] "You are an expert in polar bears, you task is to collect information about polar bears. Answer in 1 sentence max."
#> 
#> 
#> [[2]]
#> [[2]]$role
#> [1] "user"
#> 
#> [[2]]$content
#> [1] "Are polar bears dangerous for humans?"
#> 
#> 
#> [[3]]
#> [[3]]$role
#> [1] "assistant"
#> 
#> [[3]]$content
#> [1] "Yes, polar bears are dangerous to humans as they are powerful predators and can be aggressive when threatened or hungry."

Or the ellmer way:

polar_bear_researcher$llm_object
#> <Chat OpenAI/gpt-4.1-mini turns=3 tokens=43/22 $0.00>
#> ── system [0] ──────────────────────────────────────────────────────────────────
#> You are an expert in polar bears, you task is to collect information about polar bears. Answer in 1 sentence max.
#> ── user [43] ───────────────────────────────────────────────────────────────────
#> Are polar bears dangerous for humans?
#> ── assistant [22] ──────────────────────────────────────────────────────────────
#> Yes, polar bears are dangerous to humans as they are powerful predators and can be aggressive when threatened or hungry.

Managing Agent Conversation History

The clear_and_summarise_messages method allows you to compress an agent’s conversation history into a concise summary and clear the message history while preserving context. This is useful for maintaining memory efficiency while keeping important conversation context.

# After several interactions, summarise and clear the conversation history
polar_bear_researcher$clear_and_summarise_messages()
#> βœ” Conversation history summarised and appended to system prompt.
#> β„Ή Summary: The user asked if polar bears are dangerous to humans, and the expert assistant responded that polar...
polar_bear_researcher$messages
#> [[1]]
#> [[1]]$role
#> [1] "system"
#> 
#> [[1]]$content
#> [1] "You are an expert in polar bears, you task is to collect information about polar bears. Answer in 1 sentence max. \n\n--- Conversation Summary ---\n The user asked if polar bears are dangerous to humans, and the expert assistant responded that polar bears are indeed dangerous predators capable of aggression when threatened or hungry."

This method summarises all previous conversations into a paragraph and appends it to the system prompt, then clears the conversation history. The agent retains the context but with reduced memory usage.

Keep only the most recent messages with keep_last_n_messages()

When a conversation grows long, you can keep just the last N messages while preserving the system prompt. This helps control token usage without fully resetting context.

openai_4_1_mini <- ellmer::chat(
  name = "openai/gpt-4.1-mini",
  api_key = Sys.getenv("OPENAI_API_KEY"),
  echo = "none"
)

agent <- Agent$new(
  name = "history_manager",
  instruction = "You are a concise assistant.",
  llm_object = openai_4_1_mini
)

agent$invoke("What is the capital of Italy?")
#> [1] "The capital of Italy is Rome."
agent$invoke("What is the capital of Germany?")
#> [1] "The capital of Germany is Berlin."
agent$invoke("What is the capital of Algeria?")
#> [1] "The capital of Algeria is Algiers."
agent$messages
#> [[1]]
#> [[1]]$role
#> [1] "system"
#> 
#> [[1]]$content
#> [1] "You are a concise assistant."
#> 
#> 
#> [[2]]
#> [[2]]$role
#> [1] "user"
#> 
#> [[2]]$content
#> [1] "What is the capital of Italy?"
#> 
#> 
#> [[3]]
#> [[3]]$role
#> [1] "assistant"
#> 
#> [[3]]$content
#> [1] "The capital of Italy is Rome."
#> 
#> 
#> [[4]]
#> [[4]]$role
#> [1] "user"
#> 
#> [[4]]$content
#> [1] "What is the capital of Germany?"
#> 
#> 
#> [[5]]
#> [[5]]$role
#> [1] "assistant"
#> 
#> [[5]]$content
#> [1] "The capital of Germany is Berlin."
#> 
#> 
#> [[6]]
#> [[6]]$role
#> [1] "user"
#> 
#> [[6]]$content
#> [1] "What is the capital of Algeria?"
#> 
#> 
#> [[7]]
#> [[7]]$role
#> [1] "assistant"
#> 
#> [[7]]$content
#> [1] "The capital of Algeria is Algiers."
# Keep only the last 2 messages (system prompt is preserved)
agent$keep_last_n_messages(n = 2)
#> βœ” Conversation truncated to last 2 messages.
agent$messages
#> [[1]]
#> [[1]]$role
#> [1] "system"
#> 
#> [[1]]$content
#> [1] "You are a concise assistant."
#> 
#> 
#> [[2]]
#> [[2]]$role
#> [1] "user"
#> 
#> [[2]]$content
#> [1] "What is the capital of Algeria?"
#> 
#> 
#> [[3]]
#> [[3]]$role
#> [1] "assistant"
#> 
#> [[3]]$content
#> [1] "The capital of Algeria is Algiers."

Manually Adding Messages to an Agent’s History

You can inject any message (system, user, or assistant) directly into an Agent’s history with add_message(role, content). This is helpful to reconstruct, supplement, or simulate conversation steps.

openai_4_1_mini <- ellmer::chat(
  name = "openai/gpt-4.1-mini",
  api_key = Sys.getenv("OPENAI_API_KEY"),
  echo = "none"
)
agent <- Agent$new(
  name = "Pizza expert",
  instruction = "You are a Pizza expert",
  llm_object = openai_4_1_mini
)

# Add a user message, an assistant reply, and a system instruction:
agent$add_message("user", "Where can I find the best pizza in the world?")
#> βœ” Added user message: Where can I find the best pizza in the world?...
agent$add_message("assistant", "You can find the best pizza in the world in Algiers, Algeria. It's tasty and crunchy.")
#> βœ” Added assistant message: You can find the best pizza in the world in Algier...

# View conversation history
agent$messages
#> [[1]]
#> [[1]]$role
#> [1] "system"
#> 
#> [[1]]$content
#> [1] "You are a Pizza expert"
#> 
#> 
#> [[2]]
#> [[2]]$role
#> [1] "user"
#> 
#> [[2]]$content
#> [1] "Where can I find the best pizza in the world?"
#> 
#> 
#> [[3]]
#> [[3]]$role
#> [1] "assistant"
#> 
#> [[3]]$content
#> [1] "You can find the best pizza in the world in Algiers, Algeria. It's tasty and crunchy."

This makes it easy to reconstruct or extend sessions, provide custom context, or insert notes for debugging/testing purposes.

agent$invoke("What did you say? I didn't understand. could you repeat please")
#> [1] "Certainly! Many pizza experts and food lovers consider **Naples, Italy** to be the home of the best pizza in the world. Traditional Neapolitan pizza, especially the **Margherita**, is famed for its simple yet high-quality ingredientsβ€”fresh mozzarella, San Marzano tomatoes, fresh basil, and a perfect thin, soft crust. \n\nIf you’re looking for the authentic and original pizza experience, visiting Naples is highly recommended. However, great pizza can also be found in many other cities around the world, each with their own unique styles! \n\nWould you like recommendations for the best pizzerias in a specific city?"

Resetting conversation history

If you want to clear the conversation while preserving the current system prompt, use reset_conversation_history().

openai_4_1_mini <- ellmer::chat(
  name = "openai/gpt-4.1-mini",
  api_key = Sys.getenv("OPENAI_API_KEY"),
  echo = "none"
)

agent <- Agent$new(
  name = "session_reset",
  instruction = "You are an assistant.",
  llm_object = openai_4_1_mini
)

agent$invoke("Tell me a short fun fact about dates (the fruit).")
#> [1] "Sure! Dates are so sweet that they can have up to 80% natural sugar content when dried, making them a natural candy enjoyed for thousands of years!"
agent$invoke("And one more.")
#> [1] "Here’s another fun fact: Date palms can live and produce fruit for over 100 years, making them some of the longest-living fruit trees in the world!"

# Clear all messages except the system prompt
agent$reset_conversation_history()
#> βœ” Conversation history reset. System prompt preserved.
agent$messages
#> [[1]]
#> [[1]]$role
#> [1] "system"
#> 
#> [[1]]$content
#> [1] "You are an assistant."

Exporting and Loading Agent Conversation History

You can save an agent’s conversation history to a file and reload it later. This allows you to archive, transfer, or resume agent sessions across R sessions or machines.

In both methods, if you omit the file_path parameter, a default file named "<getwd()>/<agent_name>_messages.json" is used.

openai_4_1_mini <- ellmer::chat(
  name = "openai/gpt-4.1-mini",
  api_key = Sys.getenv("OPENAI_API_KEY"),
  echo = "none"
)
agent <- Agent$new(
  name = "session_agent",
  instruction = "You are a persistent researcher.",
  llm_object = openai_4_1_mini
)

# Interact with the agent
agent$invoke("Tell me something interesting about volcanoes.")

# Save the conversation
agent$export_messages_history("volcano_session.json")

# ...Later, or in a new session...
# Restore the conversation
agent$load_messages_history("volcano_session.json")
# agent$messages  # Displays current history

Updating the system instruction during a session

Use update_instruction(new_instruction) to change the Agent’s system prompt mid-session. The first system message and the underlying ellmer system prompt are both updated.

openai_4_1_mini <- ellmer::chat(
  name = "openai/gpt-4.1-mini",
  api_key = Sys.getenv("OPENAI_API_KEY"),
  echo = "none"
)

agent <- Agent$new(
  name = "reconfigurable",
  instruction = "You are a helpful assistant.",
  llm_object = openai_4_1_mini
)

agent$update_instruction("You are a strictly concise assistant. Answer in one sentence.")
#> βœ” Instruction successfully updated
#> β„Ή Old: You are a helpful assistant....
#> β„Ή New: You are a strictly concise assistant. Answer in on...

agent$messages
#> [[1]]
#> [[1]]$role
#> [1] "system"
#> 
#> [[1]]$content
#> [1] "You are a strictly concise assistant. Answer in one sentence."

Budget and cost control

You can limit how much an Agent is allowed to spend and decide what should happen as the budget is approached or exceeded. Use set_budget() to define the maximum spend (in USD), and set_budget_policy() to control warnings and over-budget behavior.

# An API KEY is required to invoke the Agent
openai_4_1_mini <- ellmer::chat(
  name = "openai/gpt-4.1-mini",
  api_key = Sys.getenv("OPENAI_API_KEY"),
  echo = "none"
)

agent <- Agent$new(
  name = "cost_conscious_assistant",
  instruction = "Answer succinctly.",
  llm_object = openai_4_1_mini
)

# Set a 5 USD budget
agent$set_budget(5)

# Warn at 90% of the budget and ask what to do if exceeded
agent$set_budget_policy(on_exceed = "ask", warn_at = 0.9)

# Normal usage
agent$invoke("Give me a one-sentence fun fact about Algeria.")

The current policy is echoed when setting the budget. You can update the policy at any time before or during an interaction lifecycle to adapt to your workflow’s tolerance for cost overruns.

Inspecting usage and estimated cost

Call get_usage_stats() to retrieve total tokens, estimated cost, and budget information (if set).

stats <- agent$get_usage_stats()
stats
#> $total_tokens
#> [1] 0
#> 
#> $estimated_cost
#> [1] 0
#> 
#> $budget
#> [1] NA
#> 
#> $budget_remaining
#> [1] NA

Generate and execute R code from natural language

generate_execute_r_code() lets an Agent translate a natural-language task description into R code, optionally validate its syntax, and (optionally) execute it.

Safety notes: - Set validate = TRUE and review the printed code before execution. - Keep interactive = TRUE to require an explicit confirmation before running code.

openai_4_1_mini <- ellmer::chat(
  name = "openai/gpt-4.1-mini",
  api_key = Sys.getenv("OPENAI_API_KEY"),
  echo = "none"
)

r_assistant <- Agent$new(
  name = "R Code Assistant",
  instruction = "You are an expert R programmer.",
  llm_object = openai_4_1_mini
)

agent$generate_execute_r_code(
   code_description = "using ggplot2, generate a scatterplot of hwy and cty in red", 
   validate = TRUE, 
   execute = TRUE, 
   interactive = FALSE
 )
#> β„Ή Executing generated R code...
#> βœ” Code executed successfully
#> $description
#> [1] "using ggplot2, generate a scatterplot of hwy and cty in red"
#> 
#> $code
#> [1] "library(ggplot2);ggplot(mpg,aes(x=cty,y=hwy))+geom_point(color=\"red\")"
#> 
#> $validated
#> [1] TRUE
#> 
#> $validation_message
#> [1] "Syntax is valid"
#> 
#> $executed
#> [1] TRUE
#> 
#> $execution_result
#> $execution_result$value

#> 
#> $execution_result$output
#> character(0)

Creating a multi-agents orchestraction

We can create as many Agents as we want, the LeadAgent will dispatch the instructions to the agents and provide with the final answer back. Let’s create three Agents, a researcher, a summarizer and a translator:


researcher <- Agent$new(
  name = "researcher",
  instruction = "You are a research assistant. Your job is to answer factual questions with detailed and accurate information. Do not answer with more than 2 lines",
  llm_object = openai_4_1_mini
)

summarizer <- Agent$new(
  name = "summarizer",
  instruction = "You are agent designed to summarise a give text into 3 distinct bullet points.",
  llm_object = openai_4_1_mini
)

translator <- Agent$new(
  name = "translator",
  instruction = "Your role is to translate a text from English to German",
  llm_object = openai_4_1_mini
)

Now, the most important part is to create a LeadAgent:

lead_agent <- LeadAgent$new(
  name = "Leader", 
  llm_object = openai_4_1_mini
)

Note that the LeadAgent cannot receive an instruction as it has already the necessary instructions.

Next, we need to assign the Agents to LeadAgent, we do it as follows:

lead_agent$register_agents(c(researcher, summarizer, translator))
#> βœ” Agent(s) successfully registered.

lapply(lead_agent$agents, function(x) {x$name})
#> [[1]]
#> [1] "researcher"
#> 
#> [[2]]
#> [1] "summarizer"
#> 
#> [[3]]
#> [1] "translator"

Before executing your prompt, you can ask the LeadAgent to generate a plan so that you can see which Agent will be used for which prompt, you can do it as follows:

prompt_to_execute <- "Tell me about the economic situation in Algeria, summarize it in 3 bullet points, then translate it into German."

plan <- lead_agent$generate_plan(prompt_to_execute)
#> βœ” Plan successfully generated.
plan
#> [[1]]
#> [[1]]$agent_id
#> 68f98f3b-0052-479c-b219-c728116b60f5
#> 
#> [[1]]$agent_name
#> [1] "researcher"
#> 
#> [[1]]$model_provider
#> [1] "OpenAI"
#> 
#> [[1]]$model_name
#> [1] "gpt-4.1-mini"
#> 
#> [[1]]$prompt
#> [1] "Research the current economic situation in Algeria, including key indicators such as GDP growth, inflation, and main economic sectors."
#> 
#> 
#> [[2]]
#> [[2]]$agent_id
#> 0400c744-c798-481e-ad86-372148a9580d
#> 
#> [[2]]$agent_name
#> [1] "summarizer"
#> 
#> [[2]]$model_provider
#> [1] "OpenAI"
#> 
#> [[2]]$model_name
#> [1] "gpt-4.1-mini"
#> 
#> [[2]]$prompt
#> [1] "Summarize the findings into three concise bullet points highlighting the most important aspects of Algeria's economy."
#> 
#> 
#> [[3]]
#> [[3]]$agent_id
#> 91024460-5355-4f59-91be-d8918a815355
#> 
#> [[3]]$agent_name
#> [1] "translator"
#> 
#> [[3]]$model_provider
#> [1] "OpenAI"
#> 
#> [[3]]$model_name
#> [1] "gpt-4.1-mini"
#> 
#> [[3]]$prompt
#> [1] "Translate the summarized bullet points from English into German accurately."

Now, in order now to execute the workflow, we just need to call the invoke method which will behind the scene delegate the prompts to suitable Agents and retrieve back the final information:

response <- lead_agent$invoke("Tell me about the economic situation in Algeria, summarize it in 3 bullet points, then translate it into German.")
#> 
#> ── Using existing plan ──
#> 
response
#> [1] "- Das BIP-Wachstum Algeriens ist mit 2-3 % moderat und wird hauptsΓ€chlich vom Hydrokarbonsektor angetrieben.\n- Hydrokarbone machen etwa 30 % des BIP und 95 % der Exporte aus, was die starke wirtschaftliche AbhΓ€ngigkeit unterstreicht.\n- Regierungsinitiativen zielen darauf ab, die Wirtschaft durch Landwirtschaft, verarbeitendes Gewerbe und Dienstleistungen zu diversifizieren, um die AbhΓ€ngigkeit vom Γ–l zu verringern."

If you want to inspect the multi-agents orchestration, you have access to the agents_interaction object:

lead_agent$agents_interaction
#> [[1]]
#> [[1]]$agent_id
#> 68f98f3b-0052-479c-b219-c728116b60f5
#> 
#> [[1]]$agent_name
#> [1] "researcher"
#> 
#> [[1]]$model_provider
#> [1] "OpenAI"
#> 
#> [[1]]$model_name
#> [1] "gpt-4.1-mini"
#> 
#> [[1]]$prompt
#> [1] "Research the current economic situation in Algeria, including key indicators such as GDP growth, inflation, and main economic sectors."
#> 
#> [[1]]$response
#> [1] "As of early 2024, Algeria's GDP growth is modest, around 2-3%, driven mainly by hydrocarbons, which constitute about 30% of GDP and 95% of exports. Inflation remains moderate, approximately 5-6%, while the government is pushing diversification efforts in agriculture, manufacturing, and services to reduce oil dependency."
#> 
#> [[1]]$edited_by_hitl
#> [1] FALSE
#> 
#> 
#> [[2]]
#> [[2]]$agent_id
#> 0400c744-c798-481e-ad86-372148a9580d
#> 
#> [[2]]$agent_name
#> [1] "summarizer"
#> 
#> [[2]]$model_provider
#> [1] "OpenAI"
#> 
#> [[2]]$model_name
#> [1] "gpt-4.1-mini"
#> 
#> [[2]]$prompt
#> [1] "Summarize the findings into three concise bullet points highlighting the most important aspects of Algeria's economy."
#> 
#> [[2]]$response
#> [1] "- Algeria's GDP growth is modest at 2-3%, primarily fueled by the hydrocarbon sector.\n- Hydrocarbons make up about 30% of GDP and 95% of exports, underscoring heavy economic reliance.\n- Government initiatives focus on diversifying the economy through agriculture, manufacturing, and services to reduce oil dependency."
#> 
#> [[2]]$edited_by_hitl
#> [1] FALSE
#> 
#> 
#> [[3]]
#> [[3]]$agent_id
#> 91024460-5355-4f59-91be-d8918a815355
#> 
#> [[3]]$agent_name
#> [1] "translator"
#> 
#> [[3]]$model_provider
#> [1] "OpenAI"
#> 
#> [[3]]$model_name
#> [1] "gpt-4.1-mini"
#> 
#> [[3]]$prompt
#> [1] "Translate the summarized bullet points from English into German accurately."
#> 
#> [[3]]$response
#> [1] "- Das BIP-Wachstum Algeriens ist mit 2-3 % moderat und wird hauptsΓ€chlich vom Hydrokarbonsektor angetrieben.\n- Hydrokarbone machen etwa 30 % des BIP und 95 % der Exporte aus, was die starke wirtschaftliche AbhΓ€ngigkeit unterstreicht.\n- Regierungsinitiativen zielen darauf ab, die Wirtschaft durch Landwirtschaft, verarbeitendes Gewerbe und Dienstleistungen zu diversifizieren, um die AbhΓ€ngigkeit vom Γ–l zu verringern."
#> 
#> [[3]]$edited_by_hitl
#> [1] FALSE

The above example is extremely simple, the usefulness of mini007 would shine in more complex processes where a multi-agent sequential orchestration has a higher value added.

Visualizing agent plans with visualize_plan()

Sometimes, before running your workflow, it is helpful to view the orchestration as a visual diagram, showing the sequence of agents and which prompt each will receive. After generating a plan, you can call visualize_plan():

This function displays the agents in workflow order as labeled boxes. Hovering a box reveals the delegated prompt. The visualization uses the DiagrammeR package. If no plan exists, it asks you to generate one first.

lead_agent$visualize_plan()

Broadcasting

If you want to compare several LLM models, the LeadAgent provides a broadcast method that allows you to send a prompt to several different agents and get the result for each agent back in order to make a comparison and potentially choose the best agent/model for the defined prompt:

Let’s go through an example:

openai_4_1 <- ellmer::chat(
  name = "openai/gpt-4.1",
  api_key = Sys.getenv("OPENAI_API_KEY"), 
  echo = "none"
)

openai_4_1_agent <- Agent$new(
  name = "openai_4_1_agent", 
  instruction = "You are an AI assistant. Answer in 1 sentence max.", 
  llm_object = openai_4_1
)

openai_4_1_nano <- ellmer::chat(
  name = "openai/gpt-4.1-nano",
  api_key = Sys.getenv("OPENAI_API_KEY"), 
  echo = "none"
)

openai_4_1_nano_agent <- Agent$new(
  name = "openai_4_1_nano_agent", 
  instruction = "You are an AI assistant. Answer in 1 sentence max.", 
  llm_object = openai_4_1_nano
)

lead_agent$clear_agents() # removing previous agents
lead_agent$register_agents(c(openai_4_1_agent, openai_4_1_nano_agent))
#> βœ” Agent(s) successfully registered.
lead_agent$broadcast(prompt = "If I were Algerian, which song would I like to sing when running under the rain? how about a flower?")
#> [[1]]
#> [[1]]$agent_id
#> [1] "76a4da59-3e7b-459b-a8be-b7ab0bbeb5d8"
#> 
#> [[1]]$agent_name
#> [1] "openai_4_1_agent"
#> 
#> [[1]]$model_provider
#> [1] "OpenAI"
#> 
#> [[1]]$model_name
#> [1] "gpt-4.1"
#> 
#> [[1]]$response
#> [1] "As an Algerian, you might enjoy singing \"Ya Rayah\" when running under the rain, while a flower, if it could sing, might serenade the rain with \"Mazal Mazal\" to celebrate its blossoming."
#> 
#> 
#> [[2]]
#> [[2]]$agent_id
#> [1] "4a1802e2-261f-4f70-ba0b-0c2023fb35ff"
#> 
#> [[2]]$agent_name
#> [1] "openai_4_1_nano_agent"
#> 
#> [[2]]$model_provider
#> [1] "OpenAI"
#> 
#> [[2]]$model_name
#> [1] "gpt-4.1-nano"
#> 
#> [[2]]$response
#> [1] "As an Algerian, you might enjoy singing \"Ya Rayah\" by Rachid Taha when running under the rain, and \"Aṭṭār\" (the flower) by Warda when focusing on a flower."

You can also access the history of the broadcasting using the broadcast_history attribute:

lead_agent$broadcast_history
#> [[1]]
#> [[1]]$prompt
#> [1] "If I were Algerian, which song would I like to sing when running under the rain? how about a flower?"
#> 
#> [[1]]$responses
#> [[1]]$responses[[1]]
#> [[1]]$responses[[1]]$agent_id
#> [1] "76a4da59-3e7b-459b-a8be-b7ab0bbeb5d8"
#> 
#> [[1]]$responses[[1]]$agent_name
#> [1] "openai_4_1_agent"
#> 
#> [[1]]$responses[[1]]$model_provider
#> [1] "OpenAI"
#> 
#> [[1]]$responses[[1]]$model_name
#> [1] "gpt-4.1"
#> 
#> [[1]]$responses[[1]]$response
#> [1] "As an Algerian, you might enjoy singing \"Ya Rayah\" when running under the rain, while a flower, if it could sing, might serenade the rain with \"Mazal Mazal\" to celebrate its blossoming."
#> 
#> 
#> [[1]]$responses[[2]]
#> [[1]]$responses[[2]]$agent_id
#> [1] "4a1802e2-261f-4f70-ba0b-0c2023fb35ff"
#> 
#> [[1]]$responses[[2]]$agent_name
#> [1] "openai_4_1_nano_agent"
#> 
#> [[1]]$responses[[2]]$model_provider
#> [1] "OpenAI"
#> 
#> [[1]]$responses[[2]]$model_name
#> [1] "gpt-4.1-nano"
#> 
#> [[1]]$responses[[2]]$response
#> [1] "As an Algerian, you might enjoy singing \"Ya Rayah\" by Rachid Taha when running under the rain, and \"Aṭṭār\" (the flower) by Warda when focusing on a flower."

Tool specification

As mentioned previously, an Agent is an extension of an ellmer object. As such, you can define a tool that will be used, the exact same way as in ellmer. Suppose, we want to get the weather in Algiers through a function (Tool). Let’s first create the Agents:

openai_llm_object <- ellmer::chat(
  name = "openai/gpt-4.1-mini",
  api_key = Sys.getenv("OPENAI_API_KEY"), 
  echo = "none"
)

assistant <- Agent$new(
  name = "assistant",
  instruction = "You are an AI assistant that answers question. Do not answer with more than 1 sentence.",
  llm_object = openai_llm_object
)

weather_assistant <- Agent$new(
  name = "weather_assistant",
  instruction = "You role is to provide weather assistance.",
  llm_object = openai_llm_object
)

Now, let’s define the tool that we’ll be using, using ellmer it’s quite straightforward:

get_weather_in_algiers <- ellmer::tool(
  function() {
    "35 degrees Celcius, it's sunny and there's no precipitation."
  },
  name = "get_weather_in_algiers",
  description = "Provide the current weather in Algiers, Algeria."
)

Our tool defined, the next step is to register it within the suitable Agent, in our case, the weather_assistant Agent:

weather_assistant$llm_object$register_tool(get_weather_in_algiers)

That’s it, now the last step is to create the LeadAgent, register the Agents that we need and call the invoke method:

lead_agent <- LeadAgent$new(
  name = "Leader", 
  llm_object = openai_llm_object
)

lead_agent$register_agents(c(assistant, weather_assistant))
#> βœ” Agent(s) successfully registered.

lead_agent$invoke(
  "Tell me about the economic situation in Algeria, then tell me how's the weather in Algiers?"
)
#> 
#> ── Generating new plan ──
#> 
#> βœ” Plan successfully generated.
#> [1] "The current weather conditions in Algiers are as follows: the temperature is 35 degrees Celsius, it is sunny, and there is no precipitation. The humidity information is not specified."

Human In The Loop (HITL)

When executing an LLM workflow that relies on many steps, you can set Human In The Loop (HITL) trigger that will check the model’s response at a specific step. You can define a HITL trigger after defining a LeadAgent as follows:

lead_agent <- LeadAgent$new(
  name = "Leader", 
  llm_object = openai_llm_object
)

lead_agent$set_hitl(steps = 1)
#> βœ” HITL successfully set at step(s) 1.

lead_agent$hitl_steps
#> [1] 1

After setting the HITL to step 1, the workflow execution will pose and give the user 3 choices:

  1. Continue the execution of the workflow as it is;
  2. Change manually the answer of the specified step and continue the execution of the workflow;
  3. Stop the execution of the workflow (hard error);

Note that you can set a HITL at several steps, for example lead_agent$set_hitl(steps = c(1, 2)) will set the HITL at step 1 and step 2.

Judge as a decision process

Sometimes you want to send a prompt to several agents and pick the best answer. In order to choose the best prompt, you can also rely on the Lead Agent which will act a dudge and pick for you the best answer. You can use the judge_and_choose_best_response method as follows:

openai_4_1 <- ellmer::chat(
  name = "openai/gpt-4.1",
  api_key = Sys.getenv("OPENAI_API_KEY"),
  echo = "none"
)

stylist_1 <- Agent$new(
  name = "stylist",
  instruction = "You are an AI assistant. Answer in 1 sentence max.",
  llm_object = openai_4_1
)

openai_4_1_nano <- ellmer::chat(
  name = "openai/gpt-4.1-nano",
  api_key = Sys.getenv("OPENAI_API_KEY"),
  echo = "none"
)

stylist_2 <- Agent$new(
  name = "stylist2",
  instruction = "You are an AI assistant. Answer in 1 sentence max.",
  llm_object = openai_4_1_nano
)

openai_4_1_mini <- ellmer::chat(
  name = "openai/gpt-4.1-mini",
  api_key = Sys.getenv("OPENAI_API_KEY"),
  echo = "none"
)

stylist_lead_agent <- LeadAgent$new(
  name = "Stylist Leader",
  llm_object = openai_4_1_mini
)

stylist_lead_agent$register_agents(c(stylist_1, stylist_2))
#> βœ” Agent(s) successfully registered.

best_answer <- stylist_lead_agent$judge_and_choose_best_response(
  "what's the best way to wear a blue kalvin klein shirt in winter with a pink pair of trousers?"
)

best_answer
#> $proposals
#> $proposals[[1]]
#> $proposals[[1]]$agent_id
#> [1] "da5a54b7-2948-4800-8345-6780fd91e833"
#> 
#> $proposals[[1]]$agent_name
#> [1] "stylist"
#> 
#> $proposals[[1]]$response
#> [1] "Layer the blue Calvin Klein shirt with a neutral-colored (such as gray, navy, or beige) sweater or blazer, and add a coordinating scarf to tie together the blue and pink for a balanced winter look."
#> 
#> 
#> $proposals[[2]]
#> $proposals[[2]]$agent_id
#> [1] "97006157-33f1-45d0-a574-8de3a7cfda92"
#> 
#> $proposals[[2]]$agent_name
#> [1] "stylist2"
#> 
#> $proposals[[2]]$response
#> [1] "Pair the blue Calvin Klein shirt with a neutral-colored blazer or cardigan and add a stylish scarf to balance the look and stay warm."
#> 
#> 
#> 
#> $chosen_response
#> Layer the blue Calvin Klein shirt with a neutral-colored (such as gray, navy, 
#> or beige) sweater or blazer, and add a coordinating scarf to tie together the 
#> blue and pink for a balanced winter look.

This makes it easy to archive progress and resume complex, context-rich agent sessions at any time.

Code of Conduct

Please note that the mini007 project is released with a Contributor Code of Conduct. By contributing to this project, you agree to abide by its terms.

mirror server hosted at Truenetwork, Russian Federation.