Agent 自动化做久了，最好给它加一层“可追责”

自动化 Agent 最怕的不是任务失败。

是它已经开始：

发邮件 改 CRM 操作财务数据 调用内部系统

但你不知道它为什么这么做。

Atom 做了一套比较少见的 Agent Governance。

Agent 一开始权限很低。

只有任务做对、结果被验证以后，才能逐渐提高 Autonomy。

同时把执行放进确定性的 Safety Layer 里。

它已经按岗位拆了 Sales、Support、Finance、Engineering 等 Agent，而且支持 Self-host。

如果准备把 Agent 真正接进公司业务，而不是只拿来聊天，这种“权限随着验证结果增加”的设计比一上来给 Full Access 稳很多。

GitHub｜rush86999/atom https://github.com/rush86999/atom?utm_source=chatgpt.com…

rush86999/atom

Source: https://github.com/rush86999/atom

What is Atom?

Atom is an open-source, self-hosted AI agent workforce — a team of specialty agents (sales, support, finance, engineering) that your people delegate to in plain language. Where other platforms sell agent capability, Atom sells agent accountability: autonomy that is earned through verified outcomes, executed inside a deterministic safety net, on your hardware.

Agents that earn trust, not assume it. Atom’s agents don’t just respond to commands — they operate autonomously within governed boundaries, handling routine work end-to-end. Every agent starts as a supervised intern and graduates through a 4-tier maturity model (STUDENT → INTERN → SUPERVISED → AUTONOMOUS) only after verified successful runs — 10/25/50 episodes, outcome-checked, not self-reported.

Verified outcomes, not self-report. Every mutating action is re-derived against your system of record by an independent postcondition oracle (on by default), and confidence is split into self-reported vs externally verified. An agent that says “done” is checked, not believed. A prompt-injected agent at any tier acts at that tier’s scoped blast radius — bounded by a default-on sandbox layer: filesystem scope, tool whitelist, tripwires, resource caps, kill-run, egress allowlist, full provenance audit. 0.027ms P99 per check.

Your data stays yours. Workflow data, agent state, and memory live on your infrastructure — embedded store, no cloud required. LLM inference uses your own API keys (BYOK, encrypted at rest) — or local models (Ollama first-class, or any local OpenAI-compatible server: LM Studio, vLLM, llama.cpp server) for fully private deployments. EU AI Act data-governance obligations (Aug 2026)? Designed for, not retrofitted.

Free edition, full features (AGPL v3): everything in this repository — every agent, integration, and governance feature — is free and open source. Keys you configure in .env are treated as BYOK and are never gated by plans or tiers. Commercial/managed editions run this same code on the client’s own infrastructure; there is no closed-source “pro” build.

💰 Budget-friendly AI agents: OpenCode Go subscription (~90% savings vs pay-per-token) — one $10/mo key unlocks general-purpose models (DeepSeek V4, Kimi K3, GLM 5.2, MiniMax M3, Qwen 3.7, Nemotron 3 Ultra, Grok 4.5) with full tool-calling & structured output — not just for coding, works for any agent workload. Setup guide →

No lock-in: 16+ LLM providers (OpenAI, Anthropic, DeepSeek, Gemini, MiniMax, Groq…) with automatic cost-aware routing, fallback, and self-healing — every run makes the next run cheaper (learning router + caching tiers).

The vision: every employee gets a personal agent team that knows their workflows, remembers context across weeks, and autonomously handles the repetitive work — research, data entry, drafting, reconciliation — while governance keeps you in control. Not AI replacing humans: AI handling the work humans shouldn’t be doing.

Receipts: 0.027ms P99 governance checks (repo benchmark) · 616k ops/s cached throughput · 69+ documented TDD hardening rounds (~1,100 fixes in the deep security sweep alone) · 85k+ test functions (84,737 across 2,759 files, verified Aug 2026). External stats sourced in docs/marketing/RESEARCH_NOTES.md; copy kit in COPY_README.md + POSITIONING.md.

💻 Quick Start

Clone to first governed workflow in ~10 minutes. Start small — one workflow, one integration, one approval gate.

git clone https://github.com/rush86999/atom.git && cd atom
make setup                 # one-shot dev bootstrap (venv, deps, .env, frontend)
make backend               # full backend on :8001
# in a second terminal:
make frontend              # Next.js UI on :3001

To use LLM features, set one key in backend/.env (or add via Settings > AI):

• OPENCODE_API_KEY for low-cost subscription coding models (~90% savings, recommended) or
• OPENAI_API_KEY / ANTHROPIC_API_KEY / DEEPSEEK_API_KEY / GOOGLE_API_KEY … or
• ATOM_LOCAL_ONLY=true + OLLAMA_BASE_URL=http://localhost:11434/v1 for fully local

Expected result: open http://localhost:3001 → sign in as [email protected] (password in backend/logs/bootstrap_admin_password.txt) → describe a workflow in plain language and watch it build with an approval gate before anything ships.

Full Quick Start → · Docker → · DigitalOcean 1-Click →

Solo operator? Skip setup — import a pre-built personal starter (invoice chase, candidate pipeline, support triage), each with an approval gate so nothing sends without your OK: Personal → Team Playbook →

🚀 AI-Generated Workflow Automation

Describe the outcome. Atom builds and runs the workflow.

Why it’s different from Zapier/Make/n8n:

• Agents, not just steps — Agents reason (not just execute): they research, decide, retry, and self-correct
• Governance built-in — Maturity gates, HITL approval, audit trail, sandbox isolation
• Self-hosted & private — Your data, your keys, your infrastructure
• Office-native — Real Excel/Word/PPTX with formula evaluation, live Canvas co-editing
• Agent-authored — You can chat with an agent to build/modify workflows (no drag-and-drop required)

Workflow Automation Guide → · Quick Start →

⚡ The AI Agent Landscape — Where Atom Fits

Atom is the only open-source platform that delivers:

• Enterprise governance (maturity tiers, HITL, audit) — without vendor lock-in
• Agents that know when to ask — autonomy earned per action type from verified track records, auto-revoked on regression; governed agent orgs with delegation contracts & privilege leases
• Self-hosted privacy — your data, your keys, your infrastructure
• Autonomous agent teammates — agents that work with your people, not just for them
• Agent-authored workflows — chat to build, no drag-and-drop
• Office/Canvas native — real Excel formulas, live co-editing
• 46+ business integrations — Salesforce, HubSpot, Slack, Jira, Stripe, QuickBooks…

📊 Comparisons

TL;DR: If you’re evaluating personal agents → Hermes/OpenClaw. If you need governed multi-agent business automation → Atom.

Full feature matrix: Why Teams Choose Atom ↓

⚡ Key Capabilities

🛡️ Production-Ready Security (Default-On)

Security Architecture → · Sandbox Deep-Dive →

📚 Documentation & Discoverability

🎯 Example Use Cases by Department

🏗️ Repository Layout

atom/
├── backend/            # FastAPI app — main_api_app:app (full) / minimal_app:app (smoke)
├── frontend-nextjs/    # Next.js web UI
├── mobile/             # React Native (Expo) companion app
├── menubar/            # Tauri macOS menubar companion
├── scripts/ · infra/ · installer/ · examples/
├── docs/               # project documentation
├── Dockerfile          # dual-app image (backend + frontend)
└── Makefile            # common tasks (start here)

🌟 Why Teams Choose Atom

Per-competitor analysis: 📊 Comparisons ↑

🤝 Contributing & Support

We welcome contributions — see CONTRIBUTING.md. Quality bar: CI-gated core suite green, typecheck clean on changed files, review required, docs updated. See docs/compliance/COMPLIANCE_MAPPING.md for the security/compliance control mapping.

• Issues: GitHub Issues
• Blog: Substack
• License: AGPL v3 — LICENSE.md