AgentLoop · ConversationLoop Python 3.13 · 自律執行中

Autonomy 自律型 AI 代理人框架

Beam Search 候選排序受管行動閘道程序技能學習為核心,
打造真正能自主感知、規劃、執行並持續進化的 AI 代理人系統。

0 內建 Procedure Skills
0 + 內建工具
0 支援 AI Provider
0 Beam Width

AgentLoop:自導向任務迴圈

每個執行回合(Turn)依固定五步驟推進,直到目標達成、被封鎖或超過步數上限

AgentLoop 是 Autonomy 的核心驅動器——一個「自導向任務迴圈」,不需要人工介入即可從目標出發,自主選擇工具、執行行動並從結果中學習。

每個 Turn 的執行流程:①從技能庫選取相關 ProcedureSkill,②由 LLM 提案 + RecipeEngine 產出候選動作路徑,③以 5 維評分排序(Beam Width=3),④透過 ActionGateway 受管執行,⑤評估結果並觸發學習。

終止條件明確:ACHIEVED(目標完成)、BLOCKED(無法繼續)、NO_CANDIDATESAPPROVAL_DENIEDMAX_STEPS_REACHED、或 FAILED——系統永遠不會陷入無限循環。

AgentLoop ActionGateway CandidateSelector LearningLoop CuratorDaemon
📚
disclose_procedure_skills() 選取相關技能
💡
generate_candidates() LLM 提案 + Recipe Engine
📊
rank_candidates() 5 維評分 · Beam Width=3
ActionGateway.execute_next() ApprovalPolicy 授權 · 受管執行
🧠
evaluate_outcome() + learn() 結果評估 · Recipe 學習

十二大技術特點

每個特點都直接對應 Autonomy 原始碼中的具體實作

自導向代理人迴圈

AgentLoop.run(goal, max_steps=12) 啟動後完全自主執行,每步驟透過 run_turn() 推進。支援互動模式(interactive=True)與批次模式,所有事件均以 run_id 記錄於 AutonomyStore

  • AgentLoop
  • RunState
  • TerminationReason

受管行動閘道

所有工具呼叫必須通過 ActionGatewayApprovalPolicy 授權。每個動作帶有 RiskLevel(LOW / MEDIUM / HIGH)、expected_effectverification_plan,確保執行可控、可審計。

  • ActionGateway
  • ApprovalPolicy
  • RiskLevel

5 維 Beam Search 評分

CandidateSelector 以 evidence_strength(+0.30)、purpose(+0.10)、risk(−0.35)、side_effects(−0.20)、penalty(−1.0)五個維度計算總分,選出前 beam_width=3 名候選路徑送往執行。

  • CandidateSelector
  • CandidatePath
  • beam_width=3

混合結果評估

DeterministicOutcomeEvaluator 優先判斷執行失敗(直接 BLOCKED),成功但結果不確定時再交由 ModelAssistedOutcomeEvaluator 呼叫 LLM 語義判斷——最小化昂貴的模型呼叫次數。

  • DeterministicEvaluator
  • ModelAssistedEvaluator
  • GoalStatus

程序技能庫(13 內建)

ProcedureSkillLibrary 內建 13 個 bundled skills,涵蓋 api-debugging、browser-navigation、code-editing、codebase-documentation、plan、procedure-skill-authoring、process-management、requesting-code-review、systematic-debugging、technical-spike、TDD、website-inspection、writing-plans,每步動態選取最相關技能載入。

  • ProcedureSkillLibrary
  • ProcedureSkill
  • requires_tools

後執行學習迴路

每次 Run 結束後,LearningLoop.review_run() 自動觸發:若結果為 ACHIEVED 且有 ≥2 個成功步驟,自動起草新 ProcedureSkillDraft(信心度 0.85),生成 LearningProposal 等待審核。

  • LearningLoop
  • LearningProposal
  • confidence=0.85

Action Recipe 引擎

RecipeEngine 監控每個成功動作的 SHA-256 指紋——當同一動作成功次數 ≥ candidate_threshold=2,自動升格為候選 Recipe,下次直接作為行動選項(source=action_skill)提出,無需 LLM 重新推理。

  • RecipeEngine
  • ActionRecipe
  • action fingerprint

技能策展守護程序

每次 Run 結束後,CuratorDaemon 在後台執行緒(daemon=True)非同步啟動 SkillCurator.apply_auto_merges(),自動偵測並合併重複(信心 0.95)或子集(信心 0.90)技能,防止技能庫膨脹。

  • CuratorDaemon
  • SkillCurator
  • async merge

可擴展工具集目錄

TOOLSET_CATALOG 涵蓋 4 個預設啟用工具集:file、terminal、search、skills,加上 opt-in 的 project(git / JSON / YAML / 測試探索)與 browser(Playwright headless Chromium,11 工具)。規劃中:memory、delegate、cronjob、computer_use。

  • ToolsetCatalog
  • ToolRegistry
  • ApprovalPolicy

完整事件溯源稽核

AutonomyStore 以 Event Sourcing 模式記錄整個執行生命週期:run_started → skills_selected → candidates_ranked → action_selected → approval_decision → observation → outcome_evaluated → recipe_learned → learning_review → run_finished,完整可重播。

  • AutonomyStore
  • record_event()
  • Event Sourcing

TUI 終端互動介面

autonomy tui 啟動豐富終端介面,提供 session 總覽面板、回合記錄、Action trail、compact/full 切換模式與 / 指令面板。自然語言直接流入 AgentLoop,UI 永遠不直接執行工具。

  • ConversationLoop
  • SessionShell
  • autonomy tui

多 Provider 支援

支援 ollama(本地)與 8 個 OpenAI-compatible providers:openai-api、nvidia、openrouter、deepseek、xai、zai、kimi-coding、alibaba。提供 API 金鑰管理與 provider 設定,autonomy doctor 可檢查端點走通性。

  • ollama
  • openrouter
  • nvidia / deepseek / xAI

5 維候選評分系統

每個候選動作路徑依以下公式計算總分,選出前 3 名送往執行

score =
evidence_strength × +0.30 purpose × +0.10 risk × −0.35 side_effects × −0.20 penalty × −1.0
evidence_strength +0.30
動作有充分依據與證據加分
purpose +0.10
有明確目的說明加分
risk −0.35
HIGH risk 扣最多,LOW 不扣分
side_effects −0.20
有副作用的動作扣分
penalty −1.0
已執行過或工具不可用,大量扣分

真實模組關係圖

每個方塊對應 Autonomy 原始碼中的實際類別或模組

Entry · Interactive TUI / CLI run
SessionShell → ConversationLoop autonomy tui · 互動式對話
AgentLoop.run(goal) autonomy run · max_steps · UUID run_id
Turn Loop · run_turn()
ProcedureSkillLibrary index() · load_selected()
RecipeEngine candidates_for() · learn()
CandidateSelector beam_width=3 · 5-dim score
Execution · ActionGateway
ApprovalPolicy authorize(action, interactive)
ActionGateway execute_next() · risk assessment
ToolRegistry execute(action) → Observation
Post-Run · finish_run()
OutcomeEvaluator Deterministic + Model-assisted
LearningLoop review_run() · draft skill
CuratorDaemon async thread · auto-merge
Persistence · AutonomyStore
📁 runs
📡 events
⚗️ recipes
📚 skills
💡 proposals
🗂️ curator_log

TOOLSET_CATALOG — 工具集 & AI Providers

fileread / write / patch / diff / outline / symbol_search / syntax_check…
terminalshell.execute / process.start / poll / log / wait / stop
skillsprocedure skill discovery & management (預設啟用)
projectgit.status/diff/log · json.parse · yaml.parse · python.test_discover (opt-in)
browserPlaywright Chromium · navigate / snapshot / click / screenshot… (opt-in)
memoryPersistent Workspace Memory & Context Injection
delegate / cronjob / computer_useplanned

支援 AI Provider

Ollama本地部署 · http://127.0.0.1:11434/v1
OpenAI APIopenai-compatible · OPENAI_API_KEY
NVIDIAintegrate.api.nvidia.com · kimi-k2.6
OpenRouterOPENROUTER_API_KEY
DeepSeekDEEPSEEK_API_KEY
xAI · Kimi · Alibaba · ZAIXAI_API_KEY · KIMI_API_KEY · DASHSCOPE_API_KEY…

13 個 Bundled Procedure Skills

每個技能以 SKILL.md 格式定義,包含名稱、描述、所需工具與平台篩選

api-debugging
browser-navigation
code-editing
codebase-documentation
plan
procedure-skill-authoring
process-management
requesting-code-review
systematic-debugging
technical-spike
test-driven-development
website-inspection
writing-plans

每次執行 Turn 時,model.select_procedure_skills() 依可用工具名稱動態篩選並載入最相關技能;執行成功且達到門檻後,LearningLoop 將自動起草新技能並生成 LearningProposal(CANDIDATE 狀態)。

AgentLoop 執行日誌

以下事件序列直接對應 AutonomyStore.record_event() 的實際呼叫順序

關於作者

Bill Liu

Bill Liu

Autonomy 框架設計者 & AI Agent 系統工程師

設計並實作了 Autonomy 自律型 AI 代理人框架,核心貢獻包括:Beam Search 候選評分系統、受管行動閘道(ActionGateway)、後執行學習迴路(LearningLoop)、以及後台技能策展守護程序(CuratorDaemon)。

致力於讓 AI 代理人在真實工程環境中可靠運行——從程式碼編輯、瀏覽器自動化到系統除錯,透過可組合的技能庫與事件溯源架構實現完整可審計性。

準備好打造你的自律型 AI 代理人了嗎?

Fork Autonomy,添加你的工具集、訓練專屬技能庫,讓代理人在你的領域自主作業。