Masai Lab

Our approach

核心理念

Three properties, treated as first-class. 三项性质，皆为一等公民。

MAS fail in ways single agents don't: cascading errors, policy drift, and runaway coordination. We design against all three from day one. 多智能体系统的失败模式与单体不同：级联错误、策略漂移、协同失控。我们从第一天就针对这三点进行设计。

Efficient

高效

Compact agent graphs, routed inference, and learned coordination protocols that cut tokens, latency, and dollars without losing capability.

紧凑的智能体图、路由化推理与可学习的协同协议，在不损失能力的前提下，降低 token、延迟与成本。

Token / task单任务 token ↓ 3–8×

Safe

安全

Formal action boundaries, policy verification, and red-teamed sandboxes — so an agent, or a group of them, never crosses lines you didn't draw.

形式化的动作边界、策略验证与红队沙箱——确保单个或一群智能体，永远不会越过你未划定的红线。

Policy violation策略违规率 < 0.1%

Resilient

韧性

Agents degrade gracefully, recover from tool outages, and survive adversarial inputs — stress-tested on failure injection, not just benchmarks.

智能体可优雅降级，从工具故障中恢复，并抵御对抗性输入——以故障注入进行压力测试，而非仅跑分。

Recovery rate故障恢复率 > 95%

The stack

技术栈

VLM · LLM · RL — composed into agents that act. VLM · LLM · RL —— 复合为能够行动的智能体。

Orchestration & Memory 编排与记忆 graph · scheduler

Reasoning · LLM 推理 · LLM plan · tool-use

Perception · VLM 感知 · VLM ground · verify

Policy learning · RL 策略学习 · RL reward · self-play

// agents = perception × reasoning × policy, composed

Our platform treats models not as endpoints but as components. Each agent carries a specific perception surface, a reasoning policy, and a learnable coordination role — orchestrated through typed messages and verifiable traces. 在我们的平台中，模型不是终点，而是组件。每一个智能体都有专属的感知面、推理策略与可学习的协同角色——通过类型化消息与可验证轨迹进行编排。

V VLM — grounding, not just vision. VLM —— 落地感知，而非仅视觉。 We use vision-language models to ground actions in the real environment and to verify what the LLM "thinks" it did. 使用视觉语言模型把动作落到真实环境，并校验 LLM "以为"自己做了什么。
L LLM — planners and critics. LLM —— 规划者与批判者。 Specialized planner, worker, and critic agents operate on structured task graphs, not freeform dialogs. 规划、执行、批判等专用智能体，作用于结构化任务图，而非自由对话。
R RL — coordination learned, not scripted. RL —— 协同是学来的，不是写死的。 Reinforcement learning tunes multi-agent policies against task reward, safety constraints, and cost budgets. 以强化学习，联合优化任务奖励、安全约束与成本预算下的多体策略。

Research directions

研究方向

Open problems we work on. 我们正在攻关的开放问题。

01 · Coordination

Learned agent protocols

可学习的智能体协议

Beyond prompt-stitching: emergent, compressed message formats learned under coordination reward, measurably reducing token overhead between agents.

超越提示拼接：在协同奖励下涌现的、可被压缩的消息格式，显著降低智能体之间的 token 开销。

02 · Safety

Multi-agent policy verification

多体策略验证

Formal and empirical techniques to bound what a group of agents can do, including collusion-style failures that no single agent could cause.

以形式化与经验方法，界定一群智能体的行为边界，涵盖单体无法制造的"共谋式"失败。

03 · Policy

RL on language-grounded agents

语言落地智能体的 RL

Fine-tuning LLM/VLM-based agents with offline and on-policy RL under tight reward budgets, without catastrophic capability loss.

在严格奖励预算下，对基于 LLM/VLM 的智能体进行离线与在线 RL 微调，且避免能力的灾难性遗忘。

04 · Resilience

Failure-aware orchestration

故障感知的编排

Scheduling that explicitly models tool outages, model fallback, and adversarial inputs — so agent groups degrade gracefully instead of silently.

显式建模工具中断、模型回退与对抗输入的调度，使智能体群优雅降级，而非悄然失效。

Building MAS in production? Let’s talk. 正在把 MAS 推向生产？来聊聊。

Masai Lab partners with research teams, infrastructure teams, and startups shipping agents into the real world. We take a small number of serious collaborations each quarter. Masai Lab 与研究团队、基础设施团队，以及将智能体推向真实世界的创业公司合作。每季度我们只接受少数认真的合作。

hello@masailab.org