First Agent

Build a working QitOS minimal coding agent from scratch: state, model, tools, run, and qita inspection.

Every QitOS agent follows one execution path: AgentModule + Engine + Trace (a trace is the structured log of decisions, actions, and observations recorded during a run). You define the strategy; the Engine owns the loop. This guide walks through that path using the minimal coding agent from examples/quickstart/minimal_agent.py.

Prerequisites

• Python 3.10+
• An API key for any OpenAI-compatible endpoint

1. Install qitos and set your model config

   pip install "qitos[models]"
   export OPENAI_API_KEY="sk-..."
   export OPENAI_BASE_URL="https://api.siliconflow.cn/v1/"
   export QITOS_MODEL="Qwen/Qwen3-8B"
   
2. Define state around a coding task

   Every minimal QitOS coding agent still starts from a typed StateSchema:

   from dataclasses import dataclass, field
   
   from qitos import StateSchema
   
   
   @dataclass
   class MinimalCodingState(StateSchema):
       scratchpad: list[str] = field(default_factory=list)
       target_file: str = "buggy_module.py"
       test_command: str = 'python -c "import buggy_module; assert buggy_module.add(20, 22) == 42"'
   

   The key mindset is that the task, scratchpad, and verification contract all live in state, not in hidden runtime glue.

3. Attach a real model and a real coding toolset

   The minimal example is intentionally model-backed. It uses an OpenAI-compatible model plus the canonical coding tool preset:

   from typing import Any
   
   from qitos import Action, AgentModule, Decision
   from qitos.kit import REACT_SYSTEM_PROMPT, ReActTextParser, format_action, render_prompt
   from qitos.kit.toolset import coding_tools
   from qitos.models import OpenAICompatibleModel
   
   
   class MinimalCodingAgent(AgentModule[MinimalCodingState, dict[str, Any], Action]):
       def __init__(self, llm: OpenAICompatibleModel, workspace_root: str) -> None:
           super().__init__(
               toolset=[coding_tools(workspace_root=workspace_root, shell_timeout=20, include_notebook=False)],
               llm=llm,
               model_parser=ReActTextParser(),
           )
   
       def init_state(self, task: str, **kwargs: Any) -> MinimalCodingState:
           return MinimalCodingState(
               task=task,
               max_steps=int(kwargs.get("max_steps", 8)),
               target_file=str(kwargs.get("target_file", "buggy_module.py")),
               test_command=str(kwargs.get("test_command")),
           )
   
       def build_system_prompt(self, state: MinimalCodingState) -> str | None:
           _ = state
           return render_prompt(
               REACT_SYSTEM_PROMPT,
               {"tool_schema": self.tool_registry.get_tool_descriptions()},
           )
   
4. Reduce trajectory back into state

   reduce() (the method that folds the latest observation and decision back into agent state) is where QitOS turns tool results into agent progress:

       def reduce(
           self,
           state: MinimalCodingState,
           observation: dict[str, Any],
           decision: Decision[Action],  # the structured output from the model
       ) -> MinimalCodingState:
           action_results = observation.get("action_results", [])
           if decision.rationale:
               state.scratchpad.append(f"Thought: {decision.rationale}")
           if decision.actions:
               state.scratchpad.append(f"Action: {format_action(decision.actions[0])}")
           if action_results:
               first = action_results[0]
               state.scratchpad.append(f"Observation: {first}")
               if isinstance(first, dict) and int(first.get("returncode", 1)) == 0:
                   state.final_result = "Patch applied and verification passed."
           return state
   

   You must return state from reduce. Forgetting to return it is still the most common mistake.

5. Seed a workspace and run the agent

   The minimal example creates one tiny buggy file, then asks the agent to fix it:

   from pathlib import Path
   
   WORKSPACE = Path("./playground/minimal_coding_agent")
   TARGET_FILE = "buggy_module.py"
   TASK = "Fix the bug in buggy_module.py and make the verification command pass."
   
   
   def seed_workspace() -> None:
       WORKSPACE.mkdir(parents=True, exist_ok=True)
       target = WORKSPACE / TARGET_FILE
       if not target.exists():
           target.write_text("def add(a, b):\n    return a - b\n", encoding="utf-8")
   
   
   def main() -> None:
       seed_workspace()
       agent = MinimalCodingAgent(llm=build_model(), workspace_root=str(WORKSPACE))
       result = agent.run(
           task=TASK,
           workspace=str(WORKSPACE),
           max_steps=8,
           target_file=TARGET_FILE,
           return_state=True,
       )
       print("final_result:", result.state.final_result)
       print("stop_reason:", result.state.stop_reason)
   
6. Inspect the run with qita

   qita board --logdir ./runs
   

   The board starts on http://127.0.0.1:8765. Use it to inspect the patching trajectory (the ordered sequence of steps the agent took), replay the run, and export the trace as HTML.

Complete example

The full minimal coding-agent example lives here:

• examples/quickstart/minimal_agent.py

You can also run the packaged version directly:

qit demo minimal

Common mistakes

Mistake	Symptom	Fix
Forget return state in reduce	StateValidationError or None state	Always end reduce() with return state
Skip provider config	OPENAI_API_KEY / QITOS_API_KEY error before the run starts	Export the API key and base URL before running the demo
Never run verification	The agent edits code but finishes without proof	Keep a concrete test_command in state and treat passing it as success
Wrong prompt-parser pairing	Parse errors every step	Match your system prompt format to ReActTextParser (a parser that extracts structured Decisions from ReAct-formatted text) or another explicit parser
Toolset too broad for the task	The run wanders through unnecessary tools	Start with the minimal coding preset and add tools only when the task really needs them

Next steps

Tutorial track

Continue with the four-lesson research path: ReAct, PlanAct, Claude Code-style, and code security audit.

Agent patterns

Compare the high-level tradeoffs between QitOS pattern families before choosing one.

Observability

Inspect every run with qita board, replay, and export.

Kit reference

Look up the parsers, planners, toolsets, and helpers used in the lessons.