Search

Build an AI Coding Agent from Scratch, Step by Step | 从零一步步搭建 AI 编程 Agent，边实现边理解 Agent 的工作原理

A security-conscious, local-first LLM agent in Go powered by Ollama & Model Context Protocol (MCP). Features defense-in-depth sandboxing, tamper-evident audit logging, and MCP tool orchestration.

具备学情诊断与持续规划能力的自适应伴学智能体（FastAPI + React + RAG，掌握度只认独立有效证据）

Desktop GUI Headless Local-first AI coding agent in Rust (axum, tokio, rmcp, portable-pty) behind HTTP + SSE, with an Angular 20 zoneless client, Tauri 2 desktop shell(and mobile in near future). OpenAI/Anthropic/Z.ai/Ollama + more, MCP bridge, permission engine, PTY terminal.

Run tool-calling LLM agents locally on Android devices using Termux and llama.cpp with complete data privacy and no external API reliance.

Per-command proxy for DeepSeek Harness: the agent decides which shell commands leave through a proxy, backed by a kernel the plugin installs and supervises itself.

Microsoft Agent Framework integration for Azure Functions.

Deterministic authorization for AI-agent tool execution

The chain-read authority port: resilient, evidence-bearing, fail-closed reads (isDelegationRevoked, verifySmartAgentSignature, isSessionDelegationAccepted) over multiple injected providers with circuit breakers, divergence detection, and bounded revocatio

Governed collaboration activities (spec 378 huddles): scope, run, participant and lifecycle contracts, admission rules as pure functions, and the media-provider port. No provider SDK, no UI.

The Developer Kit generator (spec 398 §10.3, spec 399 §3.6): `create-app <dir> --template product-repo` writes a product repository — pnpm workspace, exact @agenticprimitives/* pins, agentic.lock.json, `ap doctor` + `ap test` CI, a nightly live-gates work

The Ring-0 agentic orchestration core (ADR-0044): the intent→plan→execute→observe loop, the Planner + ToolInvoker ports, and a deterministic rule-based planner. A first-party app expresses a goal; the loop plans which tools compose to satisfy it and runs

Browser sign-in adapter seam: feature-detect FedCM and choose the browser-native path vs the guaranteed fallback. FedCM-first, not FedCM-only (ADR-0031).

The composed typed resolver (spec 346 §6.3/§8/§9): typed name → subject → derived type → (context → relationship) → permitted publication → bindings, returned as SEPARATE signals (subject / context / relationship / bindings / authority / evidence — never

Agentic Delegated Data Vault seam (spec 277) — the Vault read/write/list interface + data-classification taxonomy + persisted object envelope. Storage/runtime-agnostic; adapters (D1/R2/KMS) layer in behind it.

The two-tier context contract (spec 351 P0.11): public knowledge vs a subject's private records, entity resolution that refuses to guess, and the ContextProvider port.

The authority-aware agent harness (spec 350): the composition root that binds delegation mandates, the tool-policy risk ladder and receipts to the orchestration loop's ports.

Intent engagement protocol core (spec 336 §8): probe → response → offer → acceptance → mandate requirement → execute → receipt as pure documents, JCS digests, and two deterministic reducers. Transport-free; grants nothing.

OpenAI-compatible chat-completions binding for the @agenticprimitives/orchestration Planner / AnswerComposer ports (the in-repo adapter, chain-state-viem pattern). Vendor-neutral: the consumer supplies baseUrl + model (Groq, Ollama, OpenRouter, …). create

Apply and verify SkillRights declarations on AI agent skills: execution permitted, training rights reserved, in one line. No accounts, no telemetry, no network.