Multi-Agent Setup — Parallel Workflows and Delegation

The Hermes Agent architecture is designed to move beyond the single-loop chatbot. By leveraging its multi-agent framework, you can decompose monolithic tasks into smaller, manageable sub-tasks delegated to specialized agents. Whether you are conducting deep market research, performing a full-stack code review, or processing massive datasets, the multi-agent system allows for parallel execution and specialized expertise.

The Subagent System

At the core of Hermes’ multi-agent capabilities is the Subagent System. Instead of one agent attempting to be a generalist, a Primary Agent acts as an orchestrator (the Manager), spawning Subagents (the Specialists) to handle specific domains.

Multi-Profile Architecture

Hermes utilizes a profile-based architecture. Each agent — whether primary or subagent — is defined by a profile that includes:

System Prompt: The specific persona and constraints
Tool Access: A curated subset of tools relevant to the task
Model Configuration: The specific LLM provider (e.g., GPT-4o, Claude 3.5 Sonnet, Llama 3) and temperature

This allows you to combine the strengths of different models. For example, use a high-reasoning model (GPT-4o) as the orchestrator and a faster, cheaper model (Groq/Llama 3) for repetitive data extraction sub-tasks.

Worktree Mode: Parallelism at Scale

Worktree Mode is a specialized execution state where Hermes transforms a linear conversation into a branching tree of independent agents.

In standard mode, agents operate sequentially. In Worktree Mode, the Primary Agent can spawn multiple subagents simultaneously. Each branch of the tree operates in its own isolated context window, preventing context pollution and allowing for true parallel processing.

How Worktree Mode Operates

Root Node: The Primary Agent analyzes the global goal
Branching: The Primary Agent identifies N independent sub-tasks
Execution: N subagents are initialized concurrently
Aggregation: The Primary Agent collects outputs from all branches and synthesizes a final response

Implementing Task Delegation

Delegation in Hermes is handled via the delegate_task function. This tool allows the orchestrator to hand off a specific prompt and set of constraints to a subagent.

The `delegate_task` Signature

delegate_task({
  profileId: "researcher-pro",
  task: "Analyze the Q3 fiscal reports for the top 5 semiconductor companies.",
  expectedOutput: "A JSON array containing revenue and growth percentages.",
  timeout: 300, // seconds
  contextShare: ["market_trends.pdf", "industry_glossary.txt"]
})

Delegation Workflow

Define the Specialized Profiles

Before delegating, ensure your profiles.json contains the necessary personas. Define a coder profile with access to the filesystem and a reviewer profile with access to the linting tools.

Trigger the Orchestrator

Start your session with the Primary Agent. Provide a complex goal: “Rewrite the authentication module and ensure it passes all security audits.”

Automatic Branching

The Primary Agent invokes delegate_task for the coder to write the implementation and simultaneously invokes it for the reviewer to prepare the test suite.

Synthesis

Once the subagents return their results, the Primary Agent merges the code and the audit report into a final delivery.

Parallel Workflow Patterns

1. The Fan-Out/Fan-In Pattern

Best for Batch Processing or Comparative Research.

Fan-Out: One agent creates 10 subagents to summarize 10 different articles
Fan-In: One agent collects those 10 summaries and writes a comprehensive meta-analysis

2. The Pipeline Pattern

Best for Content Production or Software Development.

Agent A (Architect): Designs the system schema
Agent B (Developer): Implements the schema
Agent C (QA): Tests the implementation
Agent D (DevOps): Deploys the code

3. The Adversarial Pattern (Criticism Loop)

Best for Code Review or Fact Checking.

Generator Agent: Produces a draft or solution
Critic Agent: Searches for flaws, hallucinations, or bugs
Loop: The Generator refines the output based on the Critic’s feedback until a quality threshold is met

Combining Models and Providers

One of the most powerful aspects of Hermes multi-agent setup is the ability to mix and match providers within a single workflow.

Reasoning Heavy (Orchestrator) — Use Claude 3.5 Sonnet or GPT-4o. These models excel at planning, decomposition, and synthesis. They act as the Brain that decides which subagents to spawn.

Speed/Cost Optimized (Subagents) — Use Llama 3 (via Groq) or GPT-4o-mini. Ideal for high-volume tasks like data formatting, basic summarization, or regex extraction where latency is critical.

Privacy Focused (Local) — Use Mistral or Phi-3 via Ollama for subagents handling sensitive internal data that should not leave your local network.

Monitoring Subagent Progress

When running parallel workflows, visibility is key. Hermes provides a Subagent Telemetry Dashboard to monitor active branches.

Key Monitoring Metrics

Token Usage per Branch: Track which subagent is consuming the most resources
State Status: PENDING → EXECUTING → AWAITING_REVIEW → COMPLETED
Tool Call Logs: View the specific tool calls each subagent is making in real-time
Dependency Graph: A visual representation of which subagents are waiting for the output of others

If a subagent hangs or enters a loop, the Primary Agent can issue a terminate_subagent or restart_branch command to reset the state.

Real-World Use Cases

Deep Market Research

Primary Agent: Orchestrates the research project
Subagent 1 (Web Crawler): Scrapes current pricing from competitor sites
Subagent 2 (Financial Analyst): Analyzes PDF annual reports
Subagent 3 (Sentiment Analyst): Scrapes Reddit and X for user complaints
Result: A comprehensive SWOT analysis generated from three distinct data streams

Automated Code Review & Refactor

Primary Agent: Manages the PR (Pull Request)
Subagent 1 (Security Expert): Scans for SQL injection and XSS vulnerabilities
Subagent 2 (Performance Expert): Identifies O(n²) complexity issues
Subagent 3 (Style Guide Expert): Ensures adherence to PEP8 or Airbnb style guides
Result: A consolidated list of required changes with specific line-item references

Batch Content Localization

Primary Agent: Manages the localization project
Subagents (Language Specialists): Separate agents for Spanish, French, Japanese, and German, each with a profile tuned for regional idioms
Result: Parallel translation of 100+ strings with cultural nuance preserved

Next Steps and Resources

To further optimize your multi-agent deployments, explore these guides:

Advanced Prompting for Orchestrators — How to write system prompts that effectively decompose tasks
Custom Tool Integration — Creating specialized tools for your subagents
Memory Management — Managing shared state and global variables across parallel branches
API Reference — Technical documentation for delegate_task and worktree endpoints

Need help? Join the Hermes Community Discord or check the GitHub Discussions for shared profiles.json templates.