Understanding the Agent System

Agent System Overview

The Junis agent system is a set of four composable agent types — LLM, Sequential, Parallel, and Loop — that you combine into AI teams without writing code. An orchestrator routes every request to the right agent, and each agent can call tools, MCP integrations, and sub-agents. This guide explains each type, when to use it, and how to combine them.

What are the four agent types?

LLM Agent

Uses language models to generate responses, make decisions, and execute tasks

Sequential Agent

Executes agents in order, passing results from one to the next

Parallel Agent

Runs multiple agents simultaneously, combining their results

Loop Agent

Repeats agent execution until a condition is met or max iterations reached

LLM Agent

What It Does

LLM Agents are the workhorses of your AI system. They use language models (like Claude, gpt, or Gemini) to:

Generate natural language responses
Analyze and summarize text
Make decisions based on input
Call tools and APIs
Query knowledge bases (RAG)

When to Use

✅ Answering user questions ✅ Content generation (blogs, emails, summaries) ✅ Data analysis and insights ✅ Decision-making tasks ✅ Any task requiring language understanding

Configuration Options

LLM Agent Settings

name: My Agent
agent_type: LLM_AGENT
model: anthropic/claude-sonnet-5  # Or gpt, gemini, etc.

# Model Parameters
temperature: 0.7          # Creativity (0 = deterministic, 1 = creative)
max_output_tokens: 16000   # Response length limit
max_llm_calls: 3          # Max tool call iterations

# Instructions (System Prompt)
instruction: |
  You are a helpful assistant that answers questions about our product.
  Be concise, friendly, and accurate.

# Tools & Integrations
tools:
  - search_database
  - send_email
rag_tools:
  - rag_product_docs
mcp_platforms:
  - github
  - firecrawl

# Advanced
streaming_enabled: true      # Real-time response streaming
web_search_enabled: false    # Enable web search capabilities
output_key: null             # Extract specific field from response

# Extended Thinking (Claude models only)
enable_thinking: false       # Enable extended thinking mode
thinking_budget_tokens: 2048 # Token budget for thinking (min: 1024)
reasoning_effort: none       # "none", "low", "medium", "high"

Example: Customer Support Agent

Configuration
Example Conversation

name: Support Agent
agent_type: LLM_AGENT
model: anthropic/claude-sonnet-5
temperature: 0.5
instruction: |
  You are a customer support agent for Acme Corp.

  Responsibilities:
  1. Answer product questions using the knowledge base
  2. Troubleshoot common issues
  3. Escalate complex problems to human agents

  Always be empathetic and solution-oriented.

rag_tools:
  - rag_product_documentation
  - rag_faq_database

Sequential Agent

What It Does

Sequential Agents execute a series of agents in order, where each agent’s output becomes the next agent’s input. Think of it as a pipeline or assembly line.

When to Use

✅ Multi-step workflows (research → analyze → report) ✅ Data transformation pipelines ✅ Quality assurance (generate → review → refine) ✅ Complex decision trees ✅ When output of one task informs the next

Configuration Options

Sequential Agent Settings

name: My Sequential Workflow
agent_type: SEQUENTIAL_AGENT

# Sub-Agents (executed in order)
sub_agents:
  - agent_id: researcher
    order_index: 1
    description: Gathers information from multiple sources

  - agent_id: analyzer
    order_index: 2
    description: Analyzes and synthesizes research findings

  - agent_id: writer
    order_index: 3
    description: Writes final report based on analysis

# Optional: Define output format
output_key: final_report

Example: Content Creation Pipeline

Workflow Structure
Execution Flow

name: Blog Post Generator
agent_type: SEQUENTIAL_AGENT

sub_agents:
  # Step 1: Research
  - agent: Topic Researcher
    order: 1
    role: Search web and knowledge base for relevant information

  # Step 2: Outline
  - agent: Content Outliner
    order: 2
    role: Create structured outline from research

  # Step 3: Writing
  - agent: Blog Writer
    order: 3
    role: Write engaging blog post following outline

  # Step 4: SEO Optimization
  - agent: SEO Optimizer
    order: 4
    role: Add keywords, meta description, headings

Best Practices

Keep Steps Focused

Each sub-agent should have ONE clear responsibility
Don’t try to do too much in a single step
Break complex tasks into 3-5 discrete steps

Bad Example: “Research, analyze, and write report” (one agent) Good Example: Research Agent → Analysis Agent → Writing Agent (three agents)

Use Output Keys

When you need specific data from a step, use output_key to extract it:

- agent: Data Extractor
  order: 1
  output_key: extracted_data  # Only pass this field to next step

This prevents information overload in later steps.

Add Validation Steps

Insert validation agents to ensure quality:

sub_agents:
  - agent: Content Generator
    order: 1
  - agent: Quality Checker    # ← Validates output
    order: 2
  - agent: Formatter
    order: 3

Parallel Agent

What It Does

Parallel Agents run multiple agents simultaneously and combine their results. Perfect for gathering information from multiple sources or performing independent tasks concurrently.

When to Use

✅ Data collection from multiple sources ✅ Independent analyses that can run concurrently ✅ Speeding up workflows with parallelization ✅ Comparing different approaches ✅ Aggregating diverse perspectives

Configuration Options

Parallel Agent Settings

name: My Parallel Workflow
agent_type: PARALLEL_AGENT

# Sub-Agents (executed simultaneously)
sub_agents:
  - agent_id: source_a_collector
    description: Fetches data from Source A

  - agent_id: source_b_collector
    description: Fetches data from Source B

  - agent_id: source_c_collector
    description: Fetches data from Source C

# Results are automatically combined
# Order doesn't matter for parallel execution

Example: Multi-Source Data Aggregator

Workflow Structure
Execution Timeline
Combined Output

name: Market Research Aggregator
agent_type: PARALLEL_AGENT

sub_agents:
  # All run simultaneously
  - agent: Competitor Analysis Agent
    role: Analyze competitor websites and pricing

  - agent: Social Media Insights Agent
    role: Gather sentiment from Twitter, Reddit, LinkedIn

  - agent: Industry News Agent
    role: Summarize latest industry news and trends

  - agent: Customer Feedback Agent
    role: Analyze recent customer reviews and support tickets

t=0s:  All agents start simultaneously
       ├─ Competitor Analysis (15s)
       ├─ Social Media Insights (20s)
       ├─ Industry News (10s)
       └─ Customer Feedback (18s)

t=20s: All agents complete
       Results are combined into single report

Total Time: 20 seconds (vs 63 seconds if sequential)

Parallel execution requires sufficient API rate limits and resources. Monitor usage to avoid throttling.

The Parallel Agent automatically combines results:

{
  "competitor_analysis": "Our pricing is 15% lower than competitors...",
  "social_sentiment": "Positive sentiment at 68%, trending topics include...",
  "industry_news": "New regulations announced today...",
  "customer_feedback": "Top feature requests: mobile app, API access..."
}

You can add a Compiler Agent after the Parallel Agent to synthesize results into a cohesive narrative.

Best Practices

Ensure Independence

Parallel agents should NOT depend on each other’s results:Bad Example: Agent A calculates total, Agent B calculates percentage of total Good Example: Agent A analyzes Dataset 1, Agent B analyzes Dataset 2 (independent)

Handle Failures Gracefully

If one parallel agent fails, others should still complete:

# Enable fault tolerance
parallel_agent:
  continue_on_error: true  # Don't stop if one agent fails
  timeout: 30              # Max wait time per agent (seconds)

Add a Compiler Step

After parallel execution, add a Sequential step to combine results:The Compiler Agent synthesizes parallel results into a coherent response.

Loop Agent

What It Does

Loop Agents repeat execution until a condition is met or maximum iterations reached. Useful for iterative refinement, retry logic, and self-correcting workflows.

When to Use

✅ Iterative refinement (generate → evaluate → improve) ✅ Retry logic with exponential backoff ✅ Quality assurance loops ✅ Self-correcting systems ✅ Tasks requiring multiple attempts

Configuration Options

Loop Agent Settings

name: My Loop Workflow
agent_type: LOOP_AGENT
max_llm_calls: 5  # Maximum iterations (prevents infinite loops)

# Sub-Agent to Loop
sub_agents:
  - agent_id: iterative_improver
    description: Improves output each iteration

# Exit Conditions (implicit)
# - Condition evaluates to true
# - Max iterations reached
# - Agent returns success signal

Workflow Structure
Iteration Example
Max Iterations Handling

name: Code Quality Loop
agent_type: LOOP_AGENT
max_llm_calls: 3

sub_agents:
  - agent: Code Generator
    role: Generate Python function based on requirements

  - agent: Code Reviewer
    role: |
      Review code for:
      - Correctness
      - Edge cases
      - Best practices
      - Security issues

      If issues found, provide specific feedback for next iteration.
      If code is acceptable, respond with "APPROVED".

If max iterations (3) reached without approval:

{
  "status": "max_iterations_reached",
  "iterations": 3,
  "final_output": "Latest code version",
  "reviewer_feedback": "Still has issues with error handling",
  "action": "Escalate to human review"
}

Always set max_llm_calls to prevent infinite loops. Recommended: 3-5 iterations.

Best Practices

Define Clear Exit Conditions

The sub-agent should explicitly signal when to exit:

instruction: |
  Review the output. If quality is acceptable, respond with:
  "EXIT: Quality standards met"

  Otherwise, provide specific improvement suggestions.

Set Reasonable Max Iterations

2-3 iterations: Quick refinement tasks
3-5 iterations: Moderate complexity
5-10 iterations: Complex optimization (rare)

Higher values increase cost and latency. Most tasks converge in 2-3 iterations.

Track Iteration Count

Pass iteration count to the agent for context:

instruction: |
  This is iteration {{iteration_count}} of {{max_iterations}}.
  Previous feedback: {{previous_feedback}}

  Improve the output based on feedback.

Implement Degradation Strategy

Handle max iterations gracefully:

# After loop exits
if iterations >= max_llm_calls:
  escalate_to_human_review()
  log_warning("Loop did not converge")
  return best_attempt

Combining Agent Types

The real power of Junis comes from combining different agent types to build sophisticated workflows.

Example: Full Blog Post System

Orchestrator Routes Request

User says: “Write a blog post about quantum computing”Orchestrator recognizes this as a content creation task and routes to Content Pipeline (Sequential Agent).

Sequential Agent Coordinates Workflow

Content Pipeline executes three stages:

Research (Parallel Agent)
Writing (Loop Agent)
Optimization (LLM Agent)

Parallel Agent Gathers Research

Research Team runs three agents simultaneously:

Web Researcher: Latest quantum computing news
Doc Researcher: Academic papers from knowledge base
Expert Interview Agent: Queries interview transcripts

Results feed into Compiler Agent that synthesizes a research brief.

Loop Agent Refines Content

Writer + Reviewer Loop:

Writer drafts blog post from research
Reviewer checks quality, accuracy, readability
Loop continues until approved (max 3 iterations)

Final Agent Optimizes for SEO

SEO Optimizer adds:

Meta description
Keywords
Heading structure
Internal links

Output: Publication-ready blog post delivered in ~45 seconds.

Which agent type should you use?

Not sure which agent type to use? Follow this decision tree:

Decision Matrix
Complexity Guide

Scenario	Agent Type	Reason
Answer a question	LLM Agent	Single-step task
Research → Analyze → Report	Sequential	Multi-step, order matters
Check 3 databases simultaneously	Parallel	Independent, concurrent
Generate → Review → Improve	Loop	Iterative refinement
Route to appropriate handler	Orchestrator	Dynamic dispatching

Common Patterns

Pattern: Research → Analyze → Report

Use Case: Generate insights from data

Sequential Agent:
  - Parallel Research: Gather data from multiple sources
  - LLM Analysis: Identify trends and patterns
  - LLM Report Writer: Create executive summary

Pattern: Generate → Validate → Retry

Use Case: Quality assurance

Loop Agent:
  - LLM Generator: Create output
  - LLM Validator: Check against criteria
  - If invalid: Provide feedback and retry
  - If valid: Exit loop

Pattern: Triage → Route → Execute

Use Case: Smart routing

Sequential Agent:
  - LLM Triage: Classify request
  - Orchestrator: Route to specialist agent
  - Specialist LLM: Handle request

Pattern: Parallel Comparison

Use Case: Evaluate multiple approaches

Parallel Agent:
  - LLM Approach A: Conservative solution
  - LLM Approach B: Creative solution
  - LLM Approach C: Balanced solution

Then:
  - LLM Evaluator: Pick best approach

Performance Considerations

Latency

LLM Agent: 2-10 seconds
Sequential Agent: Sum of sub-agents
Parallel Agent: Max of sub-agents
Loop Agent: Iterations × agent time

Cost

LLM Agent: 1 LLM call
Sequential Agent: N LLM calls
Parallel Agent: N LLM calls (concurrent)
Loop Agent: Up to max_iterations × calls

Optimization Tips

Use Faster Models

Development: Gemini Flash (fast, cheap)
Production: Claude Sonnet (balanced)
Complex Tasks Only: Claude Opus / GPT-5.5

Switching from Opus to a faster model like Gemini 3.7 Flash can substantially reduce latency and cut per-token costs by roughly two-thirds.

Enable Streaming

streaming_enabled: true  # Show tokens as they arrive

Improves perceived performance even if total time is the same.

Limit Parallel Agents

2-3 parallel agents: Optimal for most cases
4-5 parallel agents: Acceptable with monitoring
6+ parallel agents: Risk rate limiting and high costs

Consider Sequential + Parallel hybrid instead of large Parallel groups.

Set Conservative max_llm_calls

Loop Agent:
  max_llm_calls: 3  # Rarely need more than 3 iterations

Most loops converge in 2-3 iterations. Higher values rarely improve output quality.

Next Steps

Add Tools to Agents

Connect agents to APIs and external services

MCP Integration

Integrate GitHub, Firecrawl, and custom services

Tools Overview

Let agents query your documents

Squad Collaboration

Let one squad’s orchestrator ask another squad and relay the answer

Organization Wiki

Give every agent a shared second brain that persists across conversations

Questions? Contact us at contact@junis.ai to discuss agent architectures and get expert guidance.

​Agent System Overview

​What are the four agent types?

LLM Agent

Sequential Agent

Parallel Agent

Loop Agent

​LLM Agent

​What It Does

​When to Use

​Configuration Options

​Example: Customer Support Agent

​Sequential Agent

​What It Does

​When to Use

​Configuration Options

​Example: Content Creation Pipeline

​Best Practices

​Parallel Agent

​What It Does

​When to Use

​Configuration Options

​Example: Multi-Source Data Aggregator

​Best Practices

​Loop Agent

​What It Does

​When to Use

​Configuration Options

​Example: Code Review and Refinement

​Best Practices

​Combining Agent Types

​Example: Full Blog Post System

​Which agent type should you use?

​Common Patterns

​Performance Considerations

Latency

Cost

​Optimization Tips

​Next Steps

Add Tools to Agents

MCP Integration

Tools Overview

Squad Collaboration

Organization Wiki

Agent System Overview

What are the four agent types?

LLM Agent

What It Does

When to Use

Configuration Options

Example: Customer Support Agent

Sequential Agent

What It Does

When to Use

Configuration Options

Example: Content Creation Pipeline

Best Practices

Parallel Agent

What It Does

When to Use

Configuration Options

Example: Multi-Source Data Aggregator

Best Practices

Loop Agent

What It Does

When to Use

Configuration Options

Example: Code Review and Refinement

Best Practices

Combining Agent Types

Example: Full Blog Post System

Which agent type should you use?

Common Patterns

Performance Considerations

Optimization Tips

Next Steps