Maxwell yamingmou

Hi, we're the OfferKuai Team 👋

Building production-grade AI agent systems

Products

We build products where AI agents execute on behalf of users — not suggest, not recommend, but actually do the work. Every product is a production system running thousands of agent sessions daily, with human oversight at critical decision points.

OfferKuai

Your AI job-application agent. It searches, filters, reaches out, follows up, and answers screening questions — the repetitive 80-90% of job hunting — then pings you only when a decision needs your judgment. You stay in control; the agent handles the grind.

🔗 offerkuai.com · Live · 2026

MoreYoo (in dev)

A desktop agent that learns by watching you. Show it a repetitive workflow once — data entry, report formatting, cross-system syncing — and it takes over, running on your device at your rhythm. When it's unsure, it stops and asks. What it produces looks like your work, not a robot's.

Coming soon

SmartMentra (in dev)

An AI agent for realistic interviews — two sides:

For HR: an AI avatar that conducts real interviews on your behalf — schedules calls, asks role-specific questions, evaluates answers, delivers structured reports.
For candidates: realistic mock interview practice — phone and video — with instant feedback.

zhimianshi.com · Coming soon

More products in the pipeline.

Agent Runtime Research

We study how agent harness frameworks work under the hood — scheduler loops, tool-call invariants, session persistence, subagent lifecycle handoff. Our scope spans multiple frameworks, not just one.

Deep audit

DeepSeek Harness (DSH) — Systematic source-level audit of dsh-agent-loop, dsh-subagent, dsh-tools, dsh-session. Focus: async cleanup gaps, cross-module resource lifecycle, production-grade resilience. (discussions)

Cross-framework comparison

Framework	What we study
LangGraph (LangChain)	Stateful graph execution, checkpoint persistence, scheduler architecture vs DSH
AutoGen (Microsoft)	Actor model (v0.4), session persistence, tool-call error recovery, demo-to-prod gap
CrewAI	Role-based crew orchestration, failure propagation, resource cleanup across members
Others	OpenAI Assistants API, n8n, Claude Agent SDK, Google ADK — building a cross-framework view of what production-grade agent runtime requires

🔬 Core research direction: Distributed humanized execution engine

We're exploring the architecture behind agent systems that operate on real user devices — the engine that powers our own products.

End-Cloud Synchronization

Dual-instance concurrent writes
Namespace-isolated seq
Merge on reconnect

Session-Level Locking

Lease-based single-writer guards
Auto-release on crash

Self-Healing GUI Automation

Record → validate → replay → auto-repair loop
Business-semantic validation (not just "page changed", but "changed correctly")

Circuit Breaker Pattern

Human-in-the-loop at sensitive decision points
Pause-resume semantics

Goal: an execution engine that behaves like a human operator — same IP, same device fingerprint, same rhythm — but runs 24/7 without losing context.

Tools

Tool	Description
dsh-retrace	Recall, edit-and-resend, and regenerate for DSH conversation messages
workbuddy-session-fork	Fork a conversation into an independent branch, preserving full context

How we work

Principle	What it means
Evidence first, conclusions second	We trace code paths and cite line numbers before making claims. No "probably broken" without a repro.
Architecture over symptoms	We look for patterns across modules, not one-off issues. Same invariant violation in multiple places = design gap.
Practical over theoretical	We care about what breaks in production: unbounded growth, silent corruption, permanent failures.

Currently curious about

Agent-to-agent protocols (A2A) — what does "structured delegation" look like at scale?
Session persistence invariants — what exactly must survive a crash for a conversation to be resumable?
Demo-to-prod gap — the distance between "AI can do X" and "X works 10,000 times without human intervention"

Runtime Architecture

We don't just study agent frameworks — we trace how they break. Every product above runs on runtime infrastructure we've audited, debugged, or built.

What we study & ship

Layer	Focus
Agent loop & scheduling	Single-writer leases · seq invariants · crash-safe resumption · abort handling
Tool-call protocol	`tool/call` ↔ `tool/result` pairing · async cleanup on abort/failure · tool-call timeout policy
Session persistence	Append-only logs · replay & recovery · corruption detection · format versioning
Subagent lifecycle	Cross-module ownership · disposal handoff · resource leak prevention
Distributed humanized execution	End-cloud sync · self-healing GUI automation · circuit breakers · lease-based concurrency

Runtime layers we operate at

Layer	What we trace
Framework internals	Agent loop scheduling · tool-call protocol · session persistence format · subagent lifecycle
Runtime behavior	Event loop phases · async hooks · garbage collection · process lifecycle
Protocol & serialization	WebSocket frames · JSON-RPC · append-only log formats · corruption detection

Implementation

TypeScript · Python

Infrastructure

Node.js · FastAPI · SQLite · Playwright

中文版

你好，我们是 Offer快团队 👋

构建生产级 AI Agent 系统

产品

我们构建的产品，AI Agent 真正替用户执行——不是建议、不是推荐，而是实际干活。每个产品都是生产级系统，每天运行数千个 Agent 会话，在关键决策点由人工把关。

Offer快

你的 AI 求职 Agent。它帮你搜索、筛选、打招呼、跟进、回答筛选问题——求职中 80-90% 的重复操作，只在需要你判断时才找你。你掌握决策，它干完重复活。

🔗 offerkuai.com · 已上线 · 2026

Mo悠（MoreYoo）（开发中）

一个"看一遍就会"的桌面 Agent。你在它面前演示一遍重复工作——数据录入、报表格式、跨系统搬运——它就学会了，用你的节奏在你的设备上跑。不确定时停下来问你，产出的东西像你写的。

即将推出

智面师（开发中）

拟真面试 AI Agent，服务两端：

企业端：AI 分身代替 HR 面试——自动约时间、按岗位提问、实时评估回答、输出结构化报告。
求职者端：拟真面试练习——电话 + 视频双通道，与 AI 面试官模拟真实面试场景，获得即时反馈。

zhimianshi.com · 即将推出

更多产品筹备中。

Agent 运行时研究

我们深入研究 agent harness 框架的底层机制——调度循环、工具调用不变式、会话持久化、子代理生命周期交接。研究范围覆盖多个框架。

深度审计

DeepSeek Harness (DSH) — 对 dsh-agent-loop、dsh-subagent、dsh-tools、dsh-session 进行系统性源码级审计。聚焦：异步清理缺口、跨模块资源生命周期、生产级弹性。(讨论)

跨框架对比

框架	研究方向
LangGraph（LangChain）	有状态图执行、检查点持久化、调度架构与 DSH 对比
AutoGen（Microsoft）	Actor 模型（v0.4）、会话持久化、工具调用错误恢复、Demo 到生产的鸿沟
CrewAI	角色制团队编排、失败传播、跨成员资源清理
其他	OpenAI Assistants API、n8n、Claude Agent SDK、Google ADK——建立跨框架的生产级运行时认知

🔬 核心研究方向：分布式拟人化执行引擎

我们正在探索在真实用户设备上运行的 Agent 系统架构——这也是支撑我们自己产品的引擎。

端云同步

双实例并发写
命名空间隔离的 seq
重连时合并

会话级锁

基于租约的单写入者守卫
崩溃自动释放

自愈式 GUI 自动化

录制→校验→回放→自动修复循环
业务语义校验（不只是页面变了，还要变对了）

断路器模式

敏感决策点的人工介入
暂停-恢复语义

目标：构建一个行为与真人无异（相同 IP、设备指纹、操作节奏）但 7×24 不间断运行且不丢失上下文的执行引擎。

工具

工具	描述
dsh-retrace	DSH 会话消息的撤回、编辑重发和重新生成
workbuddy-session-fork	将对话 fork 为独立分支，保留完整上下文

我们的工作方式

原则	含义
证据先行，结论后置	追踪代码路径、引用行号，再下结论。没有复现就没有"可能是坏的"。
架构优先于表象	寻找跨模块的模式，而非孤立事件。同一不变量违反在多处出现 = 设计缺口。
实践优先于理论	关注生产环境中真正会崩的东西：无界增长、静默损坏、永久性故障。

当前关注

Agent-to-Agent 协议（A2A） — 规模化场景下"结构化委派"长什么样？
会话持久化不变量 — 会话要能恢复，崩溃时必须存活的是什么？
Demo 到生产的鸿沟 — "AI 能做 X"到"X 跑一万次不需要人工干预"之间的距离

运行时架构

我们不只是研究 agent 框架——我们追踪它们怎么崩的。上面每个产品都运行在我们审计、调试或构建的运行时基础设施上。

我们研究与交付的

层级	聚焦点
Agent 循环与调度	单写入者租约 · seq 不变量 · 崩溃安全恢复 · 中止处理
工具调用协议	`tool/call` ↔ `tool/result` 配对 · abort/failure 时异步清理 · 工具调用超时策略
会话持久化	append-only 日志 · 回放与恢复 · 损坏检测 · 格式版本化
子代理生命周期	跨模块所有权 · dispose 交接 · 资源泄漏防护
分布式拟人化执行	端云同步 · 自愈式 GUI 自动化 · 断路器 · 基于租约的并发控制

我们追踪的运行时层级

层级	研究点
框架内部	Agent loop 调度 · 工具调用协议 · 会话持久化格式 · 子代理生命周期
运行时行为	事件循环阶段 · 异步 hooks · 垃圾回收 · 进程生命周期
协议与序列化	WebSocket 帧 · JSON-RPC · append-only 日志格式 · 损坏检测

实现语言