From Core to Agent

Goal

Build a practical agent by inheriting core contracts, with clear implementation milestones.

Milestone 1: define state and agent shell

1. create a StateSchema subclass with only essential fields
2. implement init_state, prepare, reduce
3. return Decision.final(...) first to verify loop plumbing

Milestone 2: wire tools and parser

1. register one simple tool
2. attach parser (for LLM output mode)
3. run one Decision.act(...) path and verify action result enters state

Milestone 3: switch to model-driven decisions

1. implement build_system_prompt
2. implement prepare
3. set decide to return None
4. verify model output can be parsed to valid Decision

Milestone 4: add env + memory/history

1. choose Env backend (HostEnv first)
2. set Engine(history_policy=...) for bounded model history
3. keep memory retrieval inside prepare(state)
4. verify trace includes memory and env payload

Milestone 5: harden runtime

1. configure budgets
2. add hook-based observability
3. add two or three regression tasks

Minimal production-ready run command

python examples/real/coding_agent.py \
  --model-base-url "https://api.siliconflow.cn/v1/" \
  --api-key "<your_api_key>" \
  --model-name "Qwen/Qwen3-8B" \
  --workspace ./playground

Source Index

• qitos/core/agent_module.py
• qitos/core/state.py
• qitos/engine/engine.py
• examples/quickstart/minimal_agent.py
• examples/real/coding_agent.py