Skip to main content
This is the simplest possible AFK agent. It demonstrates the three core concepts you need to get started: defining an Agent with a model and instructions, creating a Runner to execute it, and reading the result from final_text. If you are new to AFK, start here. Every other example builds on this foundation. 
from afk.agents import Agent
from afk.core import Runner

agent = Agent(name="chat", model="gpt-5.2-mini", instructions="Answer directly with concrete detail.")
runner = Runner()
result = runner.run_sync(agent, user_message="Define error budget in SRE.")
print(result.final_text)

​
Line-by-line explanation

Agent(...) defines the agent’s identity and behavior. The name is used for telemetry and logging. The model specifies which LLM to use. The instructions become the system prompt that guides the model’s behavior. Runner() creates the execution engine. With no arguments, it uses in-memory defaults: headless interaction mode, no telemetry sink, and no policy engine. This is the fastest way to get started during development. runner.run_sync(...) executes the agent synchronously, blocking until the run completes. Under the hood, this creates an async event loop, runs the agent through the full lifecycle (LLM call, optional tool execution, optional subagent delegation), and returns the terminal AgentResult. The user_message is the initial prompt sent to the model. result.final_text contains the model’s final text response. This is the primary output field on AgentResult. Always use final_text (not output_text) to access the agent’s response.

​
What AgentResult contains

The AgentResult dataclass returned by run_sync includes:
FieldTypeDescription
final_textstrThe agent’s final text response.
statestrTerminal state: "completed", "failed", "cancelled", or "degraded".
run_idstrUnique identifier for this run.
thread_idstrThread identifier for memory continuity across runs.
tool_executionslistRecords of all tool calls made during the run.
subagent_executionslistRecords of all subagent invocations.
usageUsageAggregateToken usage and cost estimates across all LLM calls.

​
Expected behavior

When you run this example, the runner makes a single LLM call to gpt-5.2-mini with the system instructions and user message. Since no tools are registered, the model responds with text only. The run completes in one step with state="completed", and final_text contains a concise definition of error budgets in SRE. No network calls, API keys, or external services are required beyond access to the specified LLM provider (configured via environment variables like OPENAI_API_KEY).