AI 模型和平台
Leidos 发布跨Vendor AI Agent控制系统头Way调度
在 Google 上的首选来源中添加 Unite.AI
2026年11月4日Leidos 推出 HeadWay Expot，一个治理平台，旨在监督来自不同供应商的AI agents，对其自主性设定限制，并保存对其决定的不言自明的记录。
该平台针对agents横跨不同供应商和系统的组织。 其初始框架有三部分：Know 识别活动 agents 和它们可以到达的系统；Control 确定哪些行动可以独立进行，需要批准或被禁止；以及 Prove 创建记录，允许事后对决定进行审查。 该模型将权限和证据包裹在每个agent周围，而不是仅仅将身份核实视为采取行动的充分权威。
按风险设置自主性
一份正式的HeadWay调度概况介绍解释了这些控制是如何校准的。 Agents根据任务的复杂性，后果的严重程度和业务背景，获得或失去自主权，每个行动被分配到四个结果中的一个：在采取行动前进行阻断，确认，通知和进行，或自动执行。 因此，低风险、可逆转的工作可以在不承担与相应行动相同的批准负担的情况下进行。
飞行记录员处理证据方面篡改不言而喻的推理来源和完整的监管链。 一个 Kill Switch 包含工具边界的活动，而不管云的连通性如何，而电路断路器节流电池，带宽和计算配额以限制运行中的agents. 这些功能加在一起，将许可、中断和可审计性结合起来，而不是仅仅依靠看仪表板的人。
Leidos 2026年10月9日对其治理方法的解释提供了这一设计背后的操作区别。 人行走审查要求一个人批准每个决定，而人行走监督则持续监测并升级选定的行动。 建筑限制必须支持这两种模式，在要求一个人进行干预之前适用行为限制。 对于量大的行动，有选择的升级提供了一种扩大监督范围的方法，而不会将每项行动变成核准队列中的另一个项目。
部署和操作控制
调度作为软件许可证提供，提供实施服务，并运行在客户选定的基础设施上。 其模块化设计涵盖现有的AI投资，自定义建造的agents和第三方任务agents，而不需要客户来取代他们现有的技术。 这种交付模式是供应商中立办法的实际基础：保障措施围绕着agents的组合，而基础系统则仍在。
密码和API-密钥请求通过腾讯路由，因此agents不需要直接存储这些证书。 莱多斯称这可以减少恶意演员盗取证书的风险。 运算器接收活动的统一视图，批准和提示，允许它们采取行动而不在工具之间切换。
发展工作和联邦政策
另有两个能力仍在发展之中：经验将使操作人员接口适合角色和特派团需要，情报部门将根据共同政策协调不同供应商的agents。 莱多斯还扩展了调度系统，因此当网络连接退化或消失时，其控制继续。 相关的概况介绍概念，包括角色调整接口、交叉供应商协调和重新连接后和解的地方当局，仍然是该发展议程的一部分。
莱多斯还能够协助联邦机构处理2025年4月3日发布的两份管理和预算办公室备忘录。 M-25-21涉及联邦人工智能的使用，通过作为具有法律、物质、约束性或重大影响的决定或行动的主要依据的产出，对高影响人工智能进行了定义。 就这些用途而言，需要不断进行监测、定期对操作人员进行培训和评估、人力监督、干预和问责制，并在可行时提供适当的故障保险。 这些义务为调度处的批准门槛、综合操作人员观察、干预控制和审计线索提供了对应政策。
同伴M-25-22采购备忘录涉及采购。 它要求防止供应商锁定，包括知识转让、清晰的数据和模式可移植性以及透明的许可证发放和定价。 它还要求合同条款允许各机构不断监测和评价业绩、风险和效力。 调度局选定的基础设施和多个供应商对agents的支持将产品设计与可携带性和竞争议程联系起来。
这两项备忘录都排除了作为国家安全系统组成部分的大赦国际。 M-25-22的要求不适用于"情报共同体"(Intelligence Community)元素，而M-25-21对高影响力AI的最低风险管理做法也带有同样的排除性。 这限制了OMB在国家安全和情报部署方面的比较。
发布最终是要在一个已经运作的组织agents周围放置治理界限：它们可以做些什么，何时必须停止审批，以及如何重建它们的决定。 这使得与安装的agents并存成为调度技术设计和莱多斯联邦采购论点之间的联系，而更广泛的管弦乐愿景仍然取决于仍在发展之中的能力。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
AI Models & Platforms
Leidos Unveils HeadWay Dispatch for Cross-Vendor AI Agent Control
Add Unite.AI to your preferred sources on Google
Leidos introduced HeadWay Dispatch on November 4, 2026, a governance platform designed to oversee AI agents from different vendors, set limits on their autonomy and preserve tamper-evident records of their decisions.
The platform is aimed at organizations whose agents span separate suppliers and systems. Its initial framework has three parts: Know identifies active agents and the systems they can reach; Control determines which actions can proceed independently, require approval or are prohibited; and Prove creates records that allow decisions to be reviewed afterward. The model wraps permissions and evidence around each agent instead of treating identity verification alone as sufficient authority to act.
Autonomy Set by Risk
An official HeadWay Dispatch factsheet explains how those controls are calibrated. Agents gain or lose autonomy based on task complexity, consequence severity and operational context, with each action assigned to one of four outcomes: block, confirm before acting, notify and proceed, or execute automatically. Low-risk, reversible work can therefore proceed without the same approval burden as a consequential action.
Flight Recorder handles the evidence side with tamper-evident reasoning provenance and a full chain of custody. A Kill Switch contains activity at tool boundaries regardless of cloud connectivity, while Circuit Breakers throttle battery, bandwidth and compute quotas to constrain runaway agents. Together, those features combine permissioning, interruption and auditability instead of relying solely on a person watching a dashboard.
Leidos’s October 9, 2026 explanation of its governance approach supplies the operating distinction behind that design. Human-in-the-loop review requires a person to approve every decision, while human-on-the-loop oversight monitors continuously and escalates selected actions. Architectural restrictions must underpin either model, applying behavioral limits before a person is asked to intervene. For high-volume operations, selective escalation offers a way to scale oversight without turning every action into another item in an approval queue.
Deployment and Operator Controls
Dispatch is delivered as a software license with implementation services and runs on infrastructure selected by the customer. Its modular design covers existing AI investments, custom-built agents and third-party task agents without requiring customers to replace their current technology. That delivery model is the practical basis for the vendor-neutral approach: safeguards sit around a mix of agents while the underlying systems remain in place.
Password and API-key requests are routed through Dispatch, so agents do not need to store those credentials directly. Leidos says this reduces the risk of credential theft by malicious actors. Operators receive a unified view of activity, approvals and alerts, allowing them to act without switching between tools.
Development Work and Federal Policy
Two additional capabilities remain in development: Experience would tailor operator interfaces to roles and mission needs, and Intelligence would coordinate agents from different vendors under common policies. Leidos is also extending Dispatch so its controls continue when network connections degrade or disappear. Related factsheet concepts, including role-adaptive interfaces, cross-vendor coordination and local authority with reconciliation after reconnection, remain part of that development agenda.
Leidos also says Dispatch can assist federal agencies in addressing two Office of Management and Budget memoranda issued April 3, 2025. M-25-21, which covers federal AI use, defines high-impact AI through outputs that serve as a principal basis for decisions or actions with legal, material, binding or significant effects. For those uses, it requires ongoing monitoring, periodic operator training and assessment, human oversight, intervention and accountability, plus an appropriate fail-safe where practicable. Those obligations provide a policy counterpart to Dispatch’s approval thresholds, consolidated operator view, intervention controls and audit trail.
The companion M-25-22 acquisition memorandum addresses purchasing. It calls for protections against vendor lock-in, including knowledge transfer, clear data and model portability, and transparent licensing and pricing. It also requires contract terms that allow agencies to monitor and evaluate performance, risks and effectiveness on an ongoing basis. Dispatch’s customer-selected infrastructure and support for agents from multiple suppliers connect the product design to that portability and competition agenda.
Both memoranda exclude AI used as a component of a National Security System. M-25-22’s requirements do not apply to Intelligence Community elements, and M-25-21’s minimum risk-management practices for high-impact AI carry the same exclusion. That limits how broadly the OMB comparison extends across national-security and intelligence deployments.
The launch is ultimately about placing a governance boundary around agents an organization already operates: what they may do, when they must stop for approval and how their decisions can be reconstructed. That makes coexistence with installed agents the link between Dispatch’s technical design and Leidos’s federal procurement argument, while the wider orchestration vision still depends on capabilities that remain in development.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.