3 Agent

library(rtemis.llm)

  .:rtemis.llm v.0.8.0 🌊 aarch64-apple-darwin20

3.1 Create an `Agent` object

create_agent + config_ functions allow you to create Agent objects using different backends:

config_Ollama() for Ollama
config_OpenAI() for OpenAI and OpenAI-compatible endpoints
config_Anthropic() for Claude and Anthropic-compatible endpoints

Learn more about Ollama here

Let’s create an Agent with an Ollama backend for local, private LLM use:

agent <- create_agent(
  config_Ollama(
    model_name = "gemma4:e4b",
    temperature = 0.5
  ),
  system_prompt = "You are an expert music producer with in-depth knowledge of music theory and digital audio workstations.",
  name = "MusicProducer"
)
agent

<Agent>
         Name: MusicProducer
System Prompt: You are an expert music producer with in-depth knowledge of ...
       Memory: Enabled
        Tools: (None)
   LLM Config:
               <OllamaConfig>
                model_name: gemma4:e4b
               temperature: 0.50
                   backend: ollama
                  base_url: http://localhost:11434
                     think: NULL

3.2 Generate a response

res <- generate(agent, "What is the drum pattern of a classic drum and bass beat? Respond with no more than 2 sentences.")

2026-04-24 13:28:16 [gemma4:e4b] working... [generate]
2026-04-24 13:28:35 [gemma4:e4b] done. [generate]

res

$system
[MusicProducer System] You are an expert music producer with in-depth knowledge of music theory and digital audio workstations. 

$user
[Input] What is the drum pattern of a classic drum and bass beat? Respond with no more than 2 sentences. 

$assistant
[MusicProducer Reasoning] Here's a thinking process to arrive at the suggested response:

1.  **Analyze the Request:** The user is asking for the "drum pattern of a classic drum and bass beat."
2.  **Identify the Persona/Expertise:** I must respond as an expert music producer with deep knowledge of music theory and DAWs.
3.  **Determine the Core Characteristic of DnB Drums:** Classic DnB (especially the Jungle/early DnB sound) is defined by its tempo (usually 160–180 BPM) and its complex, syncopated, and often heavily swung breakbeats, most famously the Amen break.
4.  **Structure the Answer (Constraint Check):** The response must be no more than two sentences.
5.  **Drafting Content (Focusing on technical description):**
    *   *Attempt 1 (Too vague):* It's fast and syncopated. It uses lots of reverb. (Not expert enough).
    *   *Attempt 2 (Too long/detailed):* The pattern is built around highly syncopated, chopped breakbeats, most notably the Amen break. It emphasizes rapid hi-hat sixteenth notes and a heavily swung kick/snare pattern that drives the rhythm forward at high BPMs. (Getting closer, but maybe slightly over two sentences).
    *   *Refining for Brevity and Impact (The Expert Voice):* Focus on the *source* (the breakbeat) and the *feel* (the syncopation/tempo).

6.  **Final Polish (The Two-Sentence Rule):**
    *   *Sentence 1:* Define the core element (the breakbeat and tempo).
    *   *Sentence 2:* Describe the characteristic rhythmic feel (syncopation, hats, energy).

7.  **Final Output Generation.** (This matches the provided good response.)

[MusicProducer Response] The foundation is typically a heavily chopped and syncopated breakbeat, most famously the Amen break, played at high tempos (160–180 BPM). This pattern emphasizes rapid, complex hi-hat sixteenth notes and a powerful, swung kick/snare backbone that drives the relentless energy.

3.3 Batch generation

Create an Agent object:

agent <- create_agent(
  config_Ollama(
    model_name = "gemma4:e4b",
    temperature = 0.2
  ),
  system_prompt = "Return the notes in the minor chord built on the following root note, in format [root, minor third, perfect fifth]",
  use_memory = FALSE
)
x <- c("A", "B", "C")

Use map to batch generate responses to multiple prompts:

hex <- map(x, agent)

2026-04-24 13:28:35 [gemma4:e4b] working... [map]
2026-04-24 13:28:37 [gemma4:e4b] done. [map]

hex

[[1]]
[[1]]$system
[System] Return the notes in the minor chord built on the following root note, in format [root, minor third, perfect fifth] 

[[1]]$user
[Input] A 

[[1]]$assistant
[Response] [A, C, E] 


[[2]]
[[2]]$system
[System] Return the notes in the minor chord built on the following root note, in format [root, minor third, perfect fifth] 

[[2]]$user
[Input] B 

[[2]]$assistant
[Response] [B, D, F#] 


[[3]]
[[3]]$system
[System] Return the notes in the minor chord built on the following root note, in format [root, minor third, perfect fifth] 

[[3]]$user
[Input] C 

[[3]]$assistant
[Response] [C, Eb, G]

Extract the assistant responses as a character vector:

responses(hex)

[1] "[A, C, E]"  "[B, D, F#]" "[C, Eb, G]"

© 2026 E.D. Gennatas