Sequential Workflows

A single agent can do a lot, but some problems are better solved by a chain of specialists. Sequential workflows pass the output from one agent as the input to the next — like an assembly line where each station adds value.

Why Chain Agents?

Each agent has focused instructions, a suitable model, and the right temperature for its job:

ResearchAgent (low temp, factual)
    ↓ structured facts
WriterAgent (higher temp, creative)
    ↓ polished prose

A single “do everything” agent usually produces mediocre results at every step. Specialists produce better output at each stage.

The Pattern

Orchestration is regular Ruby code — no framework or DSL required:

class ResearchAgent < RubyLLM::Agent
  model "gpt-4o"
  temperature 0.2
  instructions "You are a research assistant. Provide key facts and context."
end

class WriterAgent < RubyLLM::Agent
  model "gpt-4o"
  temperature 0.7
  instructions "You transform research notes into clear, engaging prose."
end

# Orchestration — just Ruby
topic = "How Ruby's garbage collector works"

researcher = ResearchAgent.new
research = researcher.ask("Research this topic: #{topic}").content

writer = WriterAgent.new
article = writer.ask("Write a blog post from these notes:\n\n#{research}").content

puts article

Each agent is a separate object with its own conversation history. The output of one becomes the input of the next.

Your Task

Open sequential.rb. Two agents are defined — ResearchAgent and WriterAgent. Implement the workflow:

Instantiate ResearchAgent and ask it to research the given topic
Pass the research output to WriterAgent as the basis for a blog post
Print both outputs — research notes first, then the polished article

TOPIC = "Ruby's object model and why everything is an object"

puts "--- Phase 1: Research ---"
researcher = ResearchAgent.new
research = researcher.ask("Research this topic thoroughly: #{TOPIC}")
puts research.content
puts

puts "--- Phase 2: Writing ---"
writer = WriterAgent.new
article = writer.ask(
  "Transform these research notes into a concise 2-paragraph blog post:\n\n#{research.content}"
)
puts article.content

Run it:

$ ruby agentic_workflows/sequential.rb

Adding More Stages

Extend the pipeline by adding agents:

class EditorAgent < RubyLLM::Agent
  model "gpt-4o"
  temperature 0.1
  instructions <<~PROMPT
    You are a technical editor. Review this blog post for:
    - Technical accuracy
    - Grammar and clarity
    - Appropriate length (under 300 words)
    Return the improved version directly.
  PROMPT
end

# Three-stage pipeline
research = ResearchAgent.new.ask("Research: #{topic}").content
draft = WriterAgent.new.ask("Write from:\n#{research}").content
final = EditorAgent.new.ask("Edit this:\n#{draft}").content

Keeping Context Between Stages

Sometimes you want one agent to accumulate context across stages. Use the same instance:

analyst = AnalystAgent.new

# Feed data in multiple messages
%w[sales_q1 sales_q2 sales_q3].each do |file|
  analyst.ask("Here's #{file} data: #{load(file)}")
end

# Final synthesis — agent has all three quarters in context
summary = analyst.ask("Now summarize the year-to-date trends").content

This is different from chaining — here a single agent builds up knowledge across messages.

Error Handling in Pipelines

Wrap stages in begin/rescue so one failure doesn’t kill the whole pipeline:

begin
  research = ResearchAgent.new.ask("Research: #{topic}").content
rescue RubyLLM::Error => e
  puts "Research failed: #{e.message}"
  research = "Research unavailable — write from general knowledge."
end

article = WriterAgent.new.ask("Write from:\n#{research}").content

Agents

Evaluation Loop

Built by <a href="https://chamoylabs.com">Chamoy Labs</a> based on <a href="https://rubyllm.com/">RubyLLM</a> by <a href="https://paolino.me/">Carmine Paolino</a>

Files

Preparing Environment

Preparing Ruby runtime