Skip to content

custom_agents.md

Latest commit

 

History

History
89 lines (71 loc) · 2.9 KB

custom_agents.md

File metadata and controls

89 lines (71 loc) · 2.9 KB

Custom Agents Guide

Subclassing Agent lets you add custom logic to any part of the agent workflow. Override only the methods you need for your use case.

1. Basic Subclass

from agentforge.agent import Agent

class EchoAgent(Agent):
    def build_output(self):
        self.output = self.result
  • Prompt File: .agentforge/prompts/echo_agent.yaml 
    prompts:
  system: "You echo back the user input."
  user: "Echo: {user_input}"
  • Usage: 
    agent = EchoAgent()
print(agent.run(user_input="Hello"))  # "Echo: Hello"

2. Overriding Hooks

Hook When Called Custom Use Case
process_data After loading data Clean or transform input data
parse_result After LLM response Parse JSON, extract fields
post_process_result After parsing Additional processing or side effects
build_output Last step before return Format final user-facing output

Example: JSON Parser Agent

import json
from agentforge.agent import Agent

class JSONAgent(Agent):
    def parse_result(self):
        try:
            self.parsed_result = json.loads(self.result)
        except Exception:
            self.parsed_result = {"text": self.result}
    def build_output(self):
        self.output = f"Parsed Data:\n{self.parsed_result}"
  • Prompt File: .agentforge/prompts/json_agent.yaml with a prompt asking for JSON output.

3. Switching Prompt Files

By default, AgentClass() loads a prompt file matching the class name (e.g., echo_agent.yaml). To use a different prompt file:

agent = EchoAgent(agent_name="detailed_prompt")

Ensure .agentforge/prompts/detailed_prompt.yaml exists with your desired prompts.

4. Advanced Integration

  • Inject tools or APIs by instantiating them in __init__ and using them in hooks.
  • Chain calls in process_data or parse_result to interact with other agents or services.
  • Modify self.template_data in any hook to pass dynamic variables to prompts.

5. Best Practices

  • Override only the methods you need—lean subclasses are easier to maintain.
  • Keep prompt file names in sync with your class or agent_name.
  • Use debug.mode and simulated_response in config to test logic without model calls.
  • Use self.template_data to pass all variables needed for prompt rendering.

Related Documentation

Need Help?