Contracts

Every LLM interaction in AFK flows through two typed contracts: LLMRequest (what you send) and LLMResponse (what you get back). These contracts normalize the differences between providers so your agent code never touches provider-specific types.

The flow

The adapter translates between AFK’s normalized contracts and the provider’s native format.

LLMRequest

Fields
Example

Field	Type	Purpose
`messages`	`list[Message]`	Conversation history (system, user, assistant, tool)
`model`	`str`	Model identifier
`tools`	`list[ToolSchema]`	Available tool schemas
`temperature`	`float`	Sampling temperature (0.0–2.0)
`max_tokens`	`int \| None`	Max output tokens
`top_p`	`float \| None`	Nucleus sampling
`response_format`	`ResponseFormat \| None`	Structured output format
`stop`	`list[str] \| None`	Stop sequences

from afk.llms import LLMRequest, Message

request = LLMRequest(
    messages=[
        Message(role="system", content="You are a helpful assistant."),
        Message(role="user", content="What is Python?"),
    ],
    model="gpt-5.5",
    temperature=0.7,
    max_tokens=500,
)

LLMResponse

Fields
Example

Field	Type	Purpose
`content`	`str \| None`	Text response from the model
`tool_calls`	`list[ToolCall]`	Tool calls requested by the model
`model`	`str`	Model that generated the response
`usage`	`Usage`	Token counts (prompt, completion, total)
`finish_reason`	`str`	Why generation stopped (`stop`, `tool_calls`, `length`)
`latency_ms`	`float`	End-to-end request latency

response = await client.generate(request)

print(response.content)        # "Python is a high-level programming..."
print(response.usage.total)    # 142 tokens
print(response.finish_reason)  # "stop"
print(response.latency_ms)     # 823.4

Structured output

Request structured JSON output with a Pydantic model:

from pydantic import BaseModel

class Sentiment(BaseModel):
    label: str      # "positive", "negative", "neutral"
    confidence: float
    reasoning: str

request = LLMRequest(
    messages=[Message(role="user", content="I love this product!")],
    model="gpt-5.5",
    response_format=Sentiment,   # ← Forces structured JSON output
)

response = await client.generate(request)
result = Sentiment.model_validate_json(response.content)
print(result.label)       # "positive"
print(result.confidence)  # 0.95

Message types

Role	Purpose	Source
`system`	Agent instructions	From `Agent.instructions`
`user`	User input	From `user_message` parameter
`assistant`	Model responses	Generated by the LLM
`tool`	Tool results	From tool execution

Start Here

Core Building Blocks

LLM Runtime

Production

Integrations

Contracts

The flow

LLMRequest

LLMResponse

Structured output

Message types

Next steps

Adapters

Control & Session

​The flow

​LLMRequest

​LLMResponse

​Structured output

​Message types

​Next steps

Adapters

Control & Session

The flow

LLMRequest

LLMResponse

Structured output

Message types

Next steps