Agent Workflows

Agents are type-safe, reusable LLM workflows with automatic templating and structured output support. Use agents for multi-step processes, repeated operations, or when you need guaranteed type safety across workflow boundaries.

Basic Agent

```python from patterpunk.llm.agent import Agent from patterpunk.llm.models.openai import OpenAiModel

class SummarizerAgent(Agent[str, str]): @property def model(self): return OpenAiModel(model="gpt-4")

@property def system_prompt(self): return "You are a concise summarizer. Respond with 1-2 sentences only."

@property def _user_prompt_template(self): return "Summarize this text: {{ text }}"

agent = SummarizerAgent() result = agent.execute("Long article text here...") ```

Structured Output

```python from pydantic import BaseModel

class Analysis(BaseModel): sentiment: str confidence: float topics: list[str]

class AnalyzerAgent(Agent[str, Analysis]): @property def model(self): return OpenAiModel(model="gpt-4")

@property def system_prompt(self): return "Analyze text and return structured data."

@property def _user_prompt_template(self): return "Analyze: {{ text }}"

Returns typed Analysis instance automatically

analysis = AnalyzerAgent().execute("Sample text") print(analysis.sentiment) # Type-safe access ```

Workflows

```python from patterpunk.llm.chain import AgentChain, Parallel

Sequential execution

chain = AgentChain([ ExtractorAgent(), # str -> str SummarizerAgent(), # str -> str AnalyzerAgent() # str -> Analysis ])

result = chain.execute("Input text")

Parallel execution

parallel = Parallel([ SentimentAgent(), # str -> str TopicAgent(), # str -> str KeywordAgent() # str -> str ])

Mixed workflow

complex_chain = AgentChain([ ExtractorAgent(), # Extract key content parallel, # Analyze in parallel -> list[str]

Final agent must handle list[str] input

]) ```

Practical Example

```python

Document processing pipeline

class ExtractAgent(Agent[str, str]): @property def system_prompt(self): return "Extract the main content from this document, removing headers and footers."

@property def _user_prompt_template(self): return "Document: {{ text }}"

class ClassifyAgent(Agent[str, str]): @property def system_prompt(self): return "Classify this content as: technical, business, or personal"

@property def _user_prompt_template(self): return "Content: {{ text }}"

Type-safe pipeline

pipeline = AgentChain([ExtractAgent(), ClassifyAgent()]) classification = pipeline.execute(raw_document) ```

Integration Notes

Agents work seamlessly with tools and MCP servers via the underlying Chat API. Override `execute()` method for custom logic when the default template→LLM→parse flow isn't sufficient.