TaGzxia · TalexDreamSoul

Reliable agents. Software that survives reality.

我研究并构建可验证、可回滚、由人治理的 Agent 系统，以及把这些方法带入真实生产环境的本地优先软件与 AI 原生产品。

Three bets. Everything else must earn its way back.

项目数量不再等于进展。新增投入只进入旗舰产品、生产验证和可靠 Agent 研究。

Open-source flagship

A local-first command bar and desktop agent runtime built around sandboxed plugins, typed capabilities, and a TypeScript SDK.

What makes it real

Public releases, cross-platform packaging, signed macOS delivery, plugin lifecycle and production-grade release gates.

Research program

How incidents become reusable skills, how skills decay, and how independent verifiers keep an agent from declaring success too early.

What makes it real

Longitudinal task trajectories, failure reports, governed skills and externally checkable completion evidence.

Applied systems

Multimodal products where identity, cost, payment, provider behavior and delivery provenance all have to agree.

What makes it real

Production workflows, immutable versions, human review gates, cost accounting and recoverable releases.

Working rule

I do not call a change complete because it compiles. It is complete when observable behavior, release identity and recovery evidence agree.

Observable behavior outranks a plausible patch.
Release identity must match the artifact users actually receive.
Memory and skills need scope, evidence, expiry and conflict handling.
Human review belongs at irreversible, expensive or identity-sensitive boundaries.

这些项目有真实工程价值，但不会与三条主线争夺无条件投入。

Writing from evidence

我记录程序、产品、设计与研究，但后续文章会统一遵循：问题、约束、实验、失败、证据、可复用结论。