01理念 · 思态智能Philosophy · by StateAI
传统 Agent 靠模型逐步决策。我们把任务编成明确的流程：模型负责执行，程序检查结果、控制推进和重试。Conventional agents rely on the model to decide every step. We compile the task into an explicit flow: the model executes, and the program checks results, controls progress and handles retries.
只读当前状态的局部指令，完成这一步的推理和生成。Reads only the current state's instruction and does this step's reasoning and generation.
用可执行的条件检查输出：返回码、测试、补丁是否匹配。Checks output with executable conditions: return codes, tests, whether patches apply.
记录进度和中间结果，按转移条件推进、重试或停止。Records progress and intermediate results, then advances, retries or stops by transition conditions.
为什么需要运行时Why a runtime
拖动参数，看端到端成功率怎么变化。Drag the parameters to see end-to-end success change.
示意模型，不是实验结果。逐步决策 = pⁿ；运行时每步成功率 = p·(1−(qc)^(k+1)) / (1−qc)，其中 q = 1−p。假设各步独立、重试互不相关。实测结果见第 03 部分。Illustrative model, not experimental results. Agent = pⁿ. Runtime per-step success = p·(1−(qc)^(k+1)) / (1−qc), with q = 1−p. Steps and retries are assumed independent. Measured results are in part 03.
02方案 · Agent 任务编译Solution · Agent task compiler
编译器把技能知识和控制流分开：技能知识变成每个状态内的局部指令，指导模型推理和生成；状态机记录执行进度和中间结果；明确的转移条件决定下一步。Our compiler separates skill knowledge from control flow. Knowledge becomes local instructions that guide reasoning and generation inside each state. The machine records progress and intermediate results, and explicit transition conditions decide what runs next.
03应用场景 · SWE env scalingUse case · SWE env scaling
同一套流程，分别用纯 Skill 和状态机执行。The same pipeline, run with a skill alone and with a state machine.
| 指标Metric | 纯 SkillSkill only | 状态机State machine | 变化Change |
|---|---|---|---|
| 单仓库平均耗时Avg. time per repo | 23.60min | 4.90min | −79.2% |
| 20 个仓库累计耗时Total time, 20 repos | 471.92min | 98.04min | −79.2% |
| 模型调用总数Total model calls | 818 | 58 | −92.9% |
整批生成 · 100 个实例Full batch · 100 instances
| 指标Metric | 纯 SkillSkill only | 状态机State machine | 变化Change |
|---|---|---|---|
| 可用实例数Usable instances | 75 | 94 | +25.3% |
| 有效候选数（计划 1,000）Valid candidates (1,000 planned) | 511 | 700 | +37.0% |
| 已测试候选有效率Valid rate of tested candidates | 52.41% | 70.42% | +18.0 pp |
| 每分钟产出可用实例Usable instances per minute | 0.819 | 0.921 | +12.5% |
| 每分钟产出可用补丁Usable patches per minute | 5.58 | 6.86 | +22.9% |
bug 生成 · 400 个候选Bug generation · 400 candidates
| 策略Strategy | 每边候选Candidates | 纯 SkillSkill only | 状态机State machine | 变化Change |
|---|---|---|---|---|
| 程序修改Procedural modification | 200 | 177 | 180 | +1.7% |
| LM 错误改写LM bug rewrite | 100 | 42 | 72 | +71.4% |
| LM 完整重写LM full rewrite | 100 | 29 | 57 | +96.6% |
| 合计Total | 400 | 248 | 309 | +24.6% |
| 有效率Valid rate | 62.0% | 77.3% | +15.3 pp |
04演示 · Agent as a ServiceDemo · Agent as a Service
你只管提交任务，不用搭环境、写流程、盯运行。思态智能在云端调度自己的 Agent，按编译好的状态机把任务跑完，最后直接把结果交给你：环境、任务实例，以及每一步都能复查的执行轨迹。You just submit the task. No environments to set up, no pipelines to write, no runs to babysit. StateAI runs it with our own agents in our cloud, on a compiled state machine, and hands you the results directly: environments, task instances, and an execution trace you can check step by step.
告诉我们你的仓库、工具和验收标准，我们把流程编译成状态机并以服务交付。Tell us your repositories, tools and acceptance criteria. We compile the workflow into a state machine and deliver it as a service.