惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

有赞技术团队
有赞技术团队
H
Hackread – Cybersecurity News, Data Breaches, AI and More
I
InfoQ
J
Java Code Geeks
Microsoft Security Blog
Microsoft Security Blog
G
Google Developers Blog
D
DataBreaches.Net
Recent Announcements
Recent Announcements
Microsoft Azure Blog
Microsoft Azure Blog
B
Blog RSS Feed
Y
Y Combinator Blog
博客园 - 【当耐特】
博客园 - 聂微东
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
大猫的无限游戏
大猫的无限游戏
P
Proofpoint News Feed
量子位
C
Check Point Blog
F
Fortinet All Blogs
罗磊的独立博客
Last Week in AI
Last Week in AI
GbyAI
GbyAI
L
LangChain Blog
博客园 - 司徒正美

Hacker News - Newest: "AI"

AI can't read an investor deck AI as an attorney? Student uses ChatGPT, Gemini to sue UW over alleged racial discrimination Hacking MCP Servers in AI Systems – The Rug Pull: Tool Changes After Approval GitHub - MeepCastana/KubeezCut: Free Web based video editor Can AI judge journalism? A Thiel-backed startup says yes, even if it risks chilling whistleblowers Coming soon: 10 Things That Matter in AI Right Now DARPA built an AI to fact-check enemy weapons claims What explains heterogeneity in AI adoption? When AI Meets Muscle: Context-Aware Electrical Stimulation Promises a New Way to Guide Human Movements - Department of Computer Science AI Changed How We Build. It Did Not Change What Matters. Linux rules on using AI-generated code - Copilot is OK, but humans must take 'full responsibility for the… Meta spins up AI version of Mark Zuckerberg to engage with employees Code Mode: Let Your AI Write Programs, Not Just Call Tools | TanStack Blog GitHub - Delavalom/graft: Go framework for building AI agents. Type-safe tools, multi-provider (OpenAI, Anthropic, Gemini, Bedrock), zero vendor SDKs. India's TCS tops estimates, says new AI models did not dent services demand Gen Z's fading AI hype Strong feeling: we are in a folded AI reality GitHub - machinarii/total-recall-catalog: A reference catalog of latest knowledge retrieval, memory & RAG systems GitHub - mensfeld/code-on-incus: Give each AI agent its own isolated machine with root, Docker, and systemd. Active defense detects and stops threats automatically.. Quantization, LoRA, and the 8% Problem: Benchmarking Local LLMs for Production AI Iran war: We spoke to the man making Lego-style AI videos that experts say are powerful propaganda Powell, Bessent discussed Anthropic's Mythos AI cyber threat with major U.S. banks GitHub - immartian/bellamem: Persistent belief-graph memory for AI agents. Retrieves decisive context by importance — not recency, not RAG, not /compact. recursive-mode: The Repo-Native Operating System for AI Engineering After the attack on Sam Altman's home, will AI CEO's go on the offensive? The biggest advance in AI since the LLM Opus 4.6 vs GPT 5.4 One Prompt Unity World Generation Test “AI polls” are fake polls Client Challenge Can AI be a 'child of God'? Inside Anthropic's meeting with Christian leaders
Autonomy — 自律型 AI 代理人系統 | Bill Liu
Bill Liu · 2026-06-20 · via Hacker News - Newest: "AI"

AgentLoop · ConversationLoop Python 3.13 · 自律執行中

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。

探索核心模組 查看原始碼

0 內建 Procedure Skills

0 + 內建工具

0 支援 AI Provider

0 Beam Width

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

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

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

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。

score =

evidence_strength × +0.30 purpose × +0.10 risk × −0.35 side_effects × −0.20 penalty × −1.0

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

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

xAI · Kimi · Alibaba · ZAIXAI_API_KEY · KIMI_API_KEY · DASHSCOPE_API_KEY…

api-debugging

browser-navigation

code-editing

codebase-documentation

plan

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)。

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。

LinkedIn GitHub

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。

前往 GitHub 聯絡作者