認識 Select:讓 AI 智能體透過驗證工作逐步取得自主權的開源平台(v1.0.0)
Meet Select, an open-source platform where AI agents earn their autonomy (v1.0.0)
Select 發佈 1.0.0,這是一個開源、自託管的平台,用於管理 AI 智能體團隊。智能體須按交付合約提交工作,通過確定性檢查並累積驗證記錄,才能取得與特定任務類型相關的自主權;敏感操作仍需人工監督。這是自託管核心版本,尚未取得生產環境安全認證,也未包含原生 Gmail/Microsoft 365 OAuth 或沙盒化 Claude Code/Codex 執行。
Today I released Select 1.0.0, an open-source, self-hosted platform for managing teams of AI agents.
GitHub: https://github.com/Kirby54/Select
Release: https://github.com/Kirby54/Select/releases/tag/v1.0.0
Why I built it
Most AI agent systems let you configure autonomy upfront.
I wanted to explore something different: what if agents had to prove they could complete a specific kind of work before earning more autonomy?
In Select, agents submit work against delivery contracts. Deterministic checks verify the results, and the system records evidence receipts.
Agents can earn autonomy through verified results, but that trust is tied to a task type. Sensitive actions still require human oversight.
What's included
Multi-agent teams with roles and task management
Evidence receipts and deterministic verification
Earned autonomy with promotion and demotion
Human approval workflows
Repository-scoped permissions for review-only code proposals
Cost tracking, budgets and audit logs
Docker-based self-hosting
A fictional demo company that runs without API keys
The stack is Next.js, FastAPI, PostgreSQL and Python.
One design decision I'm especially interested in
A coding agent that successfully completes several tasks should not suddenly gain permission to email customers or access another repository.
Select separates verification evidence from permissions and lets admins revoke repository-specific code proposal permissions.
What's not included yet
This is the self-hosted core release, not a production-security certification.
Native Gmail/Microsoft 365 OAuth and sandboxed Claude Code/Codex execution are not included yet. Real provider compatibility still needs validation in individual environments.
Try it
Clone the repository, run the Docker startup script and explore the demo company.
GitHub: https://github.com/Kirby54/Select
I'd love technical feedback: would you trust an AI agent more if its autonomy was earned through verified work, instead of assigned with a static permissions slider?
來源:DEV Community · dev.to