Confidential · Investor Briefing · 2026

Agentic Software
Factory V3

智能体软件工厂 V3

Autonomous software engineering, governed by design.
原生可治理的自主软件工程。
A deterministic orchestration layer that turns LLM coding agents into an auditable production system — models execute, they never govern.
确定性编排层，将 LLM 编码智能体转变为可审计的生产系统——模型只负责执行，绝不掌控决策。
current v0.13.85当前版本 v0.13.85 175 tagged releases175 个已发布版本 built as a native Paperclip pluginPaperclip 原生插件 works with any model provider适配任意模型供应商
state + event → decision
01 · The Problem01 · 问题

Agents can code.
Autonomy is what’s missing.

智能体会写代码，
缺的是自主治理。

✕Who reviews the output?

✕谁来复核产出？

A model grading its own work is not a review. Self-reported success is not evidence.

模型给自己打分不叫评审。自我报告的“成功”不是证据。

✕What happens on failure?

✕失败时会发生什么？

Silent retry loops, drifting context, unbounded cost — and nobody is notified.

静默重试循环、上下文漂移、无界成本——而且无人知晓。

✕Why trust the result?

✕凭什么信任结果？

“It works on my prompt” is not delivered software. Autonomy without governance stalls at demo scale.

提示词里跑得通 ≠ 交付的软件。缺乏治理的自主性止步于演示规模。

Every agent framework optimizes for the demo. Production software needs governance by construction.
每个智能体框架都在为演示优化；生产级软件需要的是“构造即治理”。
02 · The Solution02 · 解决方案

One deterministic brain.
Models execute — they never govern.

单一确定性大脑。
模型只执行，绝不决策。

Factory V3 is a deterministic Orchestrator shipped as a native Paperclip plugin — Paperclip provides the board, agents, workspaces, execution and logs; the Orchestrator is the sole authority for workflow, contracts, dispatch, retries and acceptance.
Factory V3 是以 Paperclip 原生插件交付的确定性编排器——Paperclip 提供看板、智能体、工作区、执行与日志；编排器则是工作流、契约、派发、重试与验收的唯一权威。
state + event  →  decision
The entire workflow policy is one pure, line-auditable function in src/orchestrator.ts. No I/O. No LLM in the decision path. Executing models are untrusted input; every state change is validated before it commits.
整套工作流策略就是 src/orchestrator.ts 中一个可逐行审计的纯函数。无 I/O，决策路径中没有 LLM。执行模型只是不受信任的输入；每次状态变更都经过验证后才会提交。
0
0
model calls in the governance path
治理路径中的模型调用次数
1
1
pure decision table — complete workflow policy
一张纯函数决策表——完整工作流策略
2
2
companies isolated on one governed host
同一受治主机上隔离运行的公司
03 · How It Works03 · 运行机制

Two human gates. Governed autonomy between them.

两道人工关卡，关卡之间是受治理的自主执行。

Intake 准入 validation 校验 Plan 规划 Architect 架构师 Plan approval 计划审批 🔒 human gate ① 🔒 人工关卡 ① BUILD — bounded worker DAG, isolated worktrees 构建——有界 Worker DAG，隔离 git 工作树 worker_flash worker_flash worker_spark Verify 验证 independent Verifier 独立验证者 Acceptance 人工验收 🔒 human gate ② 🔒 人工关卡 ② Done 完成 outcome closed 结果收尾
Intake
准入
validation
校验
↓
Plan
规划
Architect
架构师
↓
🔒 Plan approval
🔒 计划审批
human gate ①
人工关卡 ①
↓
BUILD — bounded worker DAG, isolated worktrees
构建 —— 有界 WORKER DAG，隔离 GIT 工作树
worker_flashworker_flashworker_spark
↓
Verify
验证
independent Verifier
独立验证者
↓
🔒 Acceptance
🔒 人工验收
human gate ②
人工关卡 ②
↓
Done
完成
outcome closed
结果收尾
Workers are bounded leaves. They cannot merge, deploy, ask humans, or advance milestone state.
Worker 是有界叶子节点。不能合并代码、不能部署、不能询问人类、不能推进里程碑状态。
Separation of powers. The Architect plans and integrates; an independent Verifier owns outcome proof.
权力分立。架构师负责规划与集成；独立的验证者掌握结果证明。
Bounded fleet. Default 2 concurrent workers per project, fleet ceiling 10 — throughput without chaos.
有界机队。每项目默认 2 个并发 Worker，机队上限 10——有吞吐，无混乱。
04 · Trust by Construction04 · 构建即信任

Safety that is enforced, not prompted.

安全是被强制执行的，不是被提示出来的。

✓Immutable contracts + leases

✓不可变契约 + 租约

A result is accepted only when deliverable, version, attempt and lease all match the active contract.

只有交付物、版本、尝试次数与租约全部匹配当前契约时，结果才会被接受。

✓Atomic CAS + outbox

✓原子 CAS + 发件箱

One compare-and-swap commit per transition; effects deliver idempotently or stay pending.

每次转换一次 CAS 原子提交；副作用要么幂等送达，要么保持待发。

✓Fail-closed admission

✓失败即关闭的准入

Stale, duplicate, or foreign runs are rejected before they ever reach a model.

过期、重复或外来的运行，在触达模型之前就会被拒绝。

✓One bounded retry

✓单次有界重试

After one retry, a native human interaction on the board — never an infinite loop.

重试一次后，在看板上发起原生人工交互——绝不无限循环。

✓Durable provider cooldowns

✓持久的供应商冷却

Cooldowns persist across restarts; fallback only on provable pre-tool provider failure.

冷却状态跨重启持久化；仅在可证实的工具前故障后才切换备用路由。

✓Hard isolation

✓强隔离

Per-company repos, workspaces and credentials. No shared state, no cross-company leaks.

每家公司独立的仓库、工作区与凭证。无共享状态，无跨公司泄漏。

05 · Mission Control05 · 指挥中心

The factory in your pocket.

把整座工厂装进口袋。

✓Your agent becomes the manager.

✓你的智能体就是厂长。

No new console to learn. You keep chatting with the assistant you already trust — it runs the factory for you.

无需学习新的控制台。你继续与已有的助手对话——由它替你运营工厂。

✓One topic per project.

✓每个项目一个话题。

Telegram topics mirror the board. Plan approvals, acceptance gates and incidents arrive as notifications — other messengers on demand.

Telegram 话题与看板一一对应。计划审批、验收关卡与事件以通知送达——其他通讯工具按需支持。

TELEGRAM · TOPIC: GUMA — M6
TELEGRAM · 话题：GUMA — M6
Factory✅ M6 Guma: 5 worker tasks done — verification passed, awaiting your acceptance.
Factory✅ M6 Guma：5 个 Worker 任务完成——验证通过，等待你的验收。
Factory⚠️ M7 plan ready — your approval is required to start.
Factory⚠️ M7 计划已就绪——需要你的批准才能开始。
Approved — proceed.
已批准——继续。
Paperclip kanban board
Native kanban board — every task, agent, cost and timeline, fully observable.
原生看板——每个任务、智能体、成本与时间线，完全可观测。
06 · Any Model06 · 任意模型

Any model. Zero lock-in.

任意模型，零供应商锁定。

The factory is model-agnostic by construction: provider choice lives in configuration, never in workflow authority. Route OpenAI, Anthropic, ZAI, DeepSeek or Moonshot Kimi through one governed chain.
工厂天然与模型解耦：供应商选择只存在于配置之中，绝不进入工作流权威。OpenAI、Anthropic、ZAI、DeepSeek、月之暗面 Kimi——全部经由同一条受治理的链路。
OpenAIAnthropicZAIDeepSeekMoonshot KimiOpenRouter

✓Policy-driven chains

✓策略驱动的模型链

Every role runs a primary + fallback chain per company — defined in config, validated at install.

每个角色按公司运行“主 + 备”模型链——由配置定义，安装时校验。

✓Subscription-first routing

✓订阅优先路由

Official usage probes (Kimi, Codex, ZAI) steer routing before generic cooldown timers.

官方用量探针（Kimi、Codex、ZAI）优先于通用冷却计时器驱动路由。

✓Provable failover

✓可证实的故障切换

Fallback fires only on structured, provably pre-tool provider failure — never on ambiguity.

仅在结构化、可证实的工具前故障时切换——绝不在模糊情况下切换。

07 · Proof, Not Promises07 · 证据，而非承诺

Shipping cadence is the traction.

持续交付本身就是牵引力。

175
tagged releases shipped, v0.6.0 → v0.13.85
个已发布版本，v0.6.0 → v0.13.85
7
frozen, versioned qualification protocols
份冻结的版本化资格认证协议
2
companies running on one governed host
家公司运行在同一受治主机上
▸Every release passes a frozen qualification protocol before it ships — offline-tested mechanisms, live readiness proofs, negative tests.每个版本都通过冻结的资格认证协议后才会发布——离线测试机制、实时就绪证明、负向测试。
▸Milestones close with fresh, direct outcome observation — accumulated self-reported PASS artifacts are not acceptance.里程碑以全新的直接结果观察收尾——累积的自我报告 PASS 不算验收。
▸Incidents recover autonomously with durable evidence; releases freeze into SHA-256-manifested clean-source bundles with pinned dependencies.事件凭持久证据自主恢复；发布冻结为带 SHA-256 清单的干净源码包，依赖全部锁定。
08 · Roadmap · Vision08 · 路线图 · 愿景

The operating system for
autonomous software companies.

自主软件公司的
操作系统。

v0.14 “RELIABLE OUTCOMES” — IN DEVELOPMENT
v0.14「可靠结果」——开发中
  • Outcome-bound verification: adversarial review against evidence receipts, independent of the producer.
  • 以结果为边界的验证：基于证据回执的对抗性评审，独立于产出方。
  • Durable concurrency: crash-safe work graphs, fenced retries, atomic capacity accounting.
  • 持久并发：崩溃安全的任务图、防串扰的重试、原子化的容量核算。
  • Autonomous recovery — always within enforceable authority.
  • 自主恢复——始终在可强制执行的权限范围内。
Humans own intent and acceptance. Agents own execution.
人类掌控意图与验收，智能体承担执行。
→ Ask us for a live board walkthrough
→ 欢迎预约一场看板实机演示
FACTORY V3 · CONFIDENTIAL
1 / 7
← → navigate · L language