Custom Agents

EXODUS agents use simple TOML configuration files with modular system prompts stored in separate Markdown files.

Agent Structure

exodus/agents/
├── single/          # Agent TOML configs
│   ├── triage_agent.toml
│   └── recon_agent.toml
└── prompts/         # Markdown prompt files
    ├── common.md
    └── recon_agent.md

Basic Configuration

[agent]
name = "recon_agent"
description = "Expert in reconnaissance and enumeration"
system_prompt = ["common.md", "recon_agent.md"]  # Load multiple prompt files
tools = ["core_bash"]
handoffs = ["triage_agent", "web_exploit_agent"]

[agent.llm]
model = "gemini/gemini-2.5-pro"
temperature = 0.5

System Prompt Options

Option 1: Modular Prompts (Recommended)

system_prompt = ["common.md", "recon_agent.md"]  # Loads from exodus/agents/prompts/

Option 2: Single File

system_prompt = ["recon_agent.md"]

Option 3: Inline Text

system_prompt = "You are a pentesting expert. Analyze targets."

Agent Configuration Options

  • name: Unique agent identifier
  • description: Brief description for handoffs
  • system_prompt: String or array of .md files from prompts/
  • tools: List of tool names (e.g., ["core_bash"])
  • handoffs: Other agents to transfer control to
  • max_iterations: Override global iteration limit (optional)

Per-Agent LLM Settings

[agent.llm]
model = "gemini/gemini-2.5-pro"    # Override default model
temperature = 0.7                   # Randomness (0-2)
max_tokens = 100000                 # Max response length
max_context_tokens = 700000         # Max tokens on context

Agent Handoffs

EXODUS supports dynamic agent delegation where agents can transfer control to other specialized agents during execution:

[agent]
name = "triage_agent"
description = "Routes requests to specialized agents"
system_prompt = "Analyze the user's request and delegate to the appropriate expert."
tools = []
handoffs = ["security_expert", "code_analyst", "recon_specialist"]

[agent.llm]
temperature = 0.3

How it works:

  1. Agent analyzes the task and determines if another agent is better suited
  2. Calls transfer_to_{agent_name} with a reason for the handoff
  3. Target agent receives context and continues execution
  4. Shared memory preserves conversation history across handoffs
  5. Global max_iterations prevents infinite loops across all agents

Context Management

EXODUS includes a context strategy to handle long-running missions without exceeding LLM token limits or losing critical information.

When an agent’s conversation history approaches the max_context_tokens limit, EXODUS automatically:

  1. Detects the threshold: Triggers when current tokens reach 90% of the limit.
  2. Self-Summarization: Uses the agent’s LLM to analyze the oldest 60% of the conversation.
  3. Keypoint Extraction: Generates a technical summary that preserves entities, tool results, and pending tasks.
  4. Memory Update: Replaces the old messages with a single summary message, freeing up context for new interactions.

Configuration

You can enable and tune this behavior in your settings.toml or per-agent configuration:

[llm]
# Global default (optional)
default_max_context_tokens = 700000

# Per-agent override in agent TOML
[agent.llm]
max_context_tokens = 500000
← All projects