当美国法律文书送达公司 ABC Legal 的 CTO Brandon Fuller 在今年早些时候向公司 1,100 名员工推出 Claude Enterprise 时，某种变化立刻发生了。公司各团队——包括诉讼文书送达、电子立案（eFiling）、出庭律师运营，以及市场、合规、财务等部门——开始自行构建自动化流程，没有人要求他们这么做。

Fuller 回忆道：“我们的用户真的蜂拥而至。他们看到了连接器和工具的易用性，突然之间，组织各处的人都开始自动化那些一直吞噬他们日常时间的任务。”

这是任何 CTO 都希望看到的采用效果。但 Fuller 发现还有进一步发展的机会：如果 ABC Legal 还能运营一支具备版本控制、可观测性并且始终在线的 AI Agent 队伍，会怎样？

这一愿景归根结底取决于基础设施。早期的 Agent 存放在构建者随意选择的位置，作为个人桌面上的计划任务运行。将它们迁出个人机器后，Agent 可以无人值守地运行，同时 Fuller 也能统一查看已构建的内容、成本以及它们昨晚是否正常运行。

因此，他部署了 Claude Managed Agents：统一的部署结构、共享工作区、单一的审计与计费界面，以及运行在云端而非个人笔记本电脑上的始终在线 Agent。

截至 2026 年 7 月，Fuller 与 ABC Legal 团队已追踪到：

• 已有 50 多个基于 Managed Agents 构建的 Agent 在生产环境运行
• 部分 Agent 所覆盖人工任务的成本，在尚未进行深度优化前最高降低约 50%
• 各部门约 310 名员工使用 Claude 处理日常工作

以下是他们如何做到这一点，以及从中学到的经验。

从热情到工程化：像对待软件一样对待每个 Agent

首次部署 Claude Managed Agents 时，Fuller 要求团队将每个 Agent 定义为代码。他认为，这是 Agent 最自然的形态。正如他所解释的：“Agent 本质上就是结构化文本，即提示词加配置；任何文本都可以放在代码库中，让全公司都能看到、审查并改进。”Agent 的提示词、工具列表、计划、凭证和记忆，都会放入与公司软件一同保存在 Git 代码库中的配置文件。任何 Agent 的变更都必须通过经人批准的拉取请求（pull request），因此每个 Agent 都具备版本历史、代码审查、回滚能力和审计轨迹。

他花了一周时间构建了一个包含两种模板的入门套件，并将其存放在专用 Git 代码库中。一种用于事件驱动型 Agent：一旦有事件发生，例如新任务到达或法院返回文档，便立即启动。另一种用于计划任务型 Agent：按照定时器每小时、每天或每周运行。每个 Agent 都存放在各自的文件夹中，并采用标准结构：JSON 配置文件、Markdown 格式的系统提示词、部署脚本和运维文档。将变更合并到主分支后，Agent 会自动部署。构建者无需编写软件；他们只需克隆代码库、复制入门模板、告诉 Claude Code Agent 应该做什么，便可获得 Agent 所需的一切：配置、提示词、凭证存储和记忆。

弥合技术鸿沟

Fuller 召集了公司由财务、市场、运营和开发部门成员组成的 15 人指导委员会——其中没有一人是软件开发人员——让他们克隆代码库，并使用 Claude Code 构建 Managed Agents。

目标是证明非开发人员也能自行构建生产级 Agent。如果每个 Agent 都必须通过开发团队流转，这一瓶颈将限制整个公司的推进速度。其安全性在于，他们并非在编写软件，而是在填写配置和提示词，Managed Agents 则提供运行时环境。

他说：“我得向他们解释什么是 PR。很多[非软件工程师]以为它指的是跑步，比如 PR，尽可能快地跑。现在他们已经在做拉取请求，并互相发送了。”

一周之内，全部 15 名员工都拥有了可运行的 Agent。这些构建者随后回到各自团队，培训其他人。一个月内，约 50 多个 Agent 已在 ABC Legal 各部门运行。每个 Agent 都有名称、负责人和唯一职责。

覆盖法律文书流程大多数环节的 Agent

ABC Legal 目前在法律立案流程及其相关运营的大多数阶段都部署了 Agent。

AI Code Reviewer 会审查四个代码库中的每一个拉取请求，通过多模型分析发现安全漏洞、性能回退以及被提交的凭证。工程师现在会在合并前等待它的审查结果。

EvidenceChain™ Delivery Agent 接管了一项客户经理过去需要手工完成的每周例行工作。ABC Legal 运营着一个专有网站 EvidenceChain.com，法院、原告和被告可通过该网站查询已在现场完成送达的记录，包括送达人是谁、何时尝试送达以及文书送达照片。一位客户希望持续从中提取特定记录。该 Agent 现在会针对匹配任务拉取数据库报告，使用 Managed Agent 内置浏览器获取每份 PDF，并每天将其交付至客户的 FTP 服务器。负责设置该流程的客户经理此前从未自动化过任何工作，只是向 Claude Code 描述需求，约一小时便完成构建。

eFiling Rejection Diagnoser 会在法院驳回一项提交时自动触发，读取任务详情、检查法院规则，并在大约一分钟内将诊断结果发布到 Slack；这项工作过去会耗费员工数小时。一个任务验证 Agent 会根据法院信息核查每一项传入任务。它通过浏览器访问法院网站，确认听证会或案件是否已正确立案、是否确实会在指定日期发生，随后根据发现结果调整任务，并标记司法辖区、法院和诉讼时效期限。

Attorney Coverage Agent 会协调律师网络以确保听证会有人出席：检查律师可用性、发送邮件，并读取有关可用时间和报价的回复，以便协调员确认覆盖安排。

在财务部门，AR-remittance Agent 会解析付款通知邮件，生成 NetSuite 付款应用文件，并将其发布到 Slack 供一键审批，之后再导入系统；另有一个每日运行的 Agent，会针对每张工程工单给出应资本化还是应费用化的结论。市场部门运行 Google Ads 分析 Agent，每周向渠道负责人发布建议。在运营部门，一个名为 Charvis 的审查 Agent 会检查已完成的送达任务，目前其判断与合规团队的一致率约为 98%。

Service-Overdue-Nudger 负责处理 ABC Legal 运营积压工作中的一级环节，即原本需要人工完成的重复性初步处理，并起草分层的每日外联消息，供人工审批。

让 Agent 更聪明：采集、调优、重复

ABC Legal 的 Agent 在人工监督下工作，将它们已完成的事项或建议发布到 Slack，员工可在线程中回复或以 emoji 作出反应。

Hank 是一名内部代码审查 Agent，会将每次审查发布到共享 Slack 频道。每条记录都会列出拉取请求及其得出的统计结果，使 Agent 决策的轨迹公开且可搜索。

Fuller 认为，所有这些反应数据都是被浪费的训练信号。但并非每个 Agent 都需要这一信号。大部分 Agent 是单任务执行器，其输出无人评分，且独立工作。对于确实收集评分反馈的 Agent，ABC Legal 采用三角色架构：独立的 Agent 共享同一个工作区、环境和凭证库，但按不同计划运行。这一模式会将 Slack 中的消息转化为经过版本控制和人工批准的 Agent 变更：

1. Initial Agent（初始 Agent） 执行工作，通常在任务到达或文档返回时实时运行，并记录每个操作的审计轨迹。
2. Harvester（采集 Agent） 每小时或每天运行，从 Slack 收集人工反馈；这些反馈以线程回复和 emoji 反应的形式到达，每一项都会成为带标签的数据点。
3. Tuner（调优 Agent） 每周运行，综合查看所有信息，并提出对提示词或配置的修改建议，而非调整模型权重。它只负责起草；由人工审查并合并拉取请求。

在 ABC Legal 的自我改进 Agent 闭环中，初始 Agent 实时执行工作，Harvester 按小时频率收集 Slack 中的人工反馈，而每周运行的 Tuner 则以拉取请求形式提出提示词和配置变更。Agent 通过开发人员已经在使用的同一套工作流实现改进。

一个例子是“deliveries-as-code”，即 Fuller 用于调优工作路由方式的 Agent 系统，最初始于 ABC Legal 旗下 50 人规模的姐妹公司 Docketly。Docketly 围绕“交付”组织工作，每项交付都有自己的路由和处理规则集。约 145 个规则集全部以 Git 中的单一 YAML 文件存在，而不是管理界面里的记录；因此，调优一项交付意味着编辑一个文件并创建拉取请求。

四个 Agent 构成了这个闭环：其中一个每周将结论发布到 Slack；Harvester 根据人工反馈将反应转化为标签；Tuner 在 YAML 上创建拉取请求；第四个 Agent 则将已合并的配置推送到生产数据库。第四个 Agent 只执行已经过人工审查和批准的内容。在实际操作中，一个标记交付路由错误的 emoji 反应，可在一周内变成该交付路由规则的一次已合并变更。审查是这个闭环中唯一的人工步骤。

为什么选择 Claude Managed Agents

在最终选择 Claude Managed Agents 作为组织的 Agent 开发框架前，Fuller 评估了多个框架。他的标准非常明确：平台必须具备版本控制、可观测会话、工作区计费、模型选择、记忆原语、MCP 连接能力，以及最关键的一点——无需维护基础设施。

平台的责任划分与 Fuller 希望的运营模式高度契合。Anthropic 的托管基础设施负责 Agent 运行所需的一切：执行循环、会话、记忆、控制台和模型本身。ABC Legal 则负责提示词、工具列表、触发逻辑、审计轨迹和结果反馈闭环。

一些能力在规模化运行中尤其重要：

• **版本控制：**每次推送都会创建新的 Agent 版本，并使用乐观锁定。回滚非常简单。
• **模型灵活性：**大多数 Agent 默认使用 Claude Sonnet；高吞吐量和快速任务使用 Claude Haiku；当更深层的推理值得额外成本时使用 Claude Opus。切换模型只需改动一行。
• **MCP 连接和凭证库：**Agent 可连接 ABC Legal 自有平台（提供超过 100 种工具）、用于报告的 Metabase、用于人机协作的 Slack，以及用于项目管理的 Atlassian。
• **计划部署：**周期性 Agent 通过 Bitbucket Pipelines 按 cron 计划运行，后者已负责处理代码库访问、密钥和计费。

ABC Legal 会追踪 AI 支出的每一美元，并按供应商、工具、团队和使用场景拆分。随着 Agent 队伍在春季逐步上线，支出持续上升；但在使用量继续增长的同时，支出于 7 月开始下降。这是下文所述效率优化工作的结果：众多 Agent 所覆盖任务的成本降低约 50%，同时各部门约 310 名员工在使用 Claude。

该公司对成本采取审慎策略：将支出导向回报可衡量的垂直运营工具和 Agent，同时保持横向聊天和构思类应用的广泛使用，并控制其成本。大多数 Agent 起初都有人类参与决策：Agent 查看任务或工单后提出建议，由人工审核后才会执行任何操作。建议要么被存储在任务中，并以横幅形式呈现，使员工可在工作流程中接受或拒绝；要么被发布到 Slack 频道，员工可在线程中回复。这些回应会构成包含正确与错误判断的标注数据集，输入 Harvester 和 Tuner 闭环，使团队能够编写评估并在前沿模型之间对 Agent 进行基准测试。一旦某个 Agent 证明在特定任务上与人工同样出色或更出色，它就会转入自动化模式，自主执行；之后它仍处于相同的度量框架中，以监测性能变化。

ABC Legal 跟踪的指标是效率比率，即 Agent 交付的价值相对于其运行成本。每个 Managed Agent 每次运行都会以工时和金额的形式，将自身价值回传到数据仓库。Agent 遵循 J 曲线：刚开始时往往处于亏损状态，因为它们较新且使用更大的模型；随后，随着团队编写评估、迁移至更便宜更快的模型并削减 token，用益会转为正值。

部署 Agent 队伍的最佳实践

Fuller 在部署 AI——尤其是 Claude Managed Agents——的经验中，总结出几项使用这项技术的工作原则：

• **将一切视为代码。**他说：“代码只是结构化文本。LLM 是文本引擎。你能将越多业务转化为代码库中的文本，Agent 就能带来越大的杠杆效应。”这既适用于传统软件，也同样适用于提示词、模式、调度规则、通知模板和业务配置。

• **从人类参与决策开始。**每个 Agent 都先发布建议供人工审查。只有在持续证明其判断与人工决策一致后，它才获得独立行动的权利。“每个 Agent 都必须在独自行动前赢得信任。它不会一开始就拥有这种信任。”

• 将 PR 用作控制界面。“如果你希望 Agent 参与一项决策，就让这项决策看起来像一个拉取请求。”逐行评论、审批工作流和不可变的审计轨迹，都可通过版本控制免费获得，并且能够自然地与

When Brandon Fuller, CTO of ABC Legal, a U.S.-based legal document delivery company, rolled out Claude Enterprise to the company's 1,100 employees earlier this year, something clicked immediately. Teams across the company (service of process, eFiling, and appearance counsel operations, plus marketing, compliance, finance, and more) started building automations on their own, without being asked.

"Our users really flocked to it," Fuller recalls. "They saw the ease of use of connectors and tools, and suddenly we had people all over the organization automating the tasks that had always eaten up their day."

It was exactly the kind of adoption any CTO hopes for. But Fuller saw an opportunity to go further: what if ABC Legal could also run a fleet of AI agents that were versioned, observable, and always on?

That ambition came down to infrastructure. Early agents lived wherever their builder happened to put them, as scheduled tasks on individual desktops. Moving them off personal machines would let them run unattended and give Fuller a single view of what had been built, what it cost, and whether it ran last night.

So he deployed Claude Managed Agents: one common deployment structure, shared workspaces, a single audit and billing surface, and always-on agents in the cloud instead of on a person’s laptop.

As of July 2026, Fuller and his team at ABC Legal have tracked:

• 50+ agents built with Managed Agents in production
• Up to ~50% reduction in the cost of the human tasks some agents cover, before heavy optimization
• ~310 employees across every department using Claude for daily work

Here’s how they got there and what they learned in the process.

From enthusiasm to engineering: treating every agent like software

When they first deployed Claude Managed Agents, Fuller had the team define every agent as code. He believes this is the natural form for an agent to take. As he explains, “an agent is really just structured text, a prompt plus configuration, and anything that is text can live in a repository where the whole company can see it, review it, and improve it.” An agent's prompt, tool list, schedule, credentials, and memory all go into configuration files kept in a git repository alongside the company's software. Nothing about an agent changes except through a pull request someone approves, which gives every agent version history, code review, rollback, and an audit trail.

He spent a week building a starter kit with two templates, stored in dedicated git repositories. One is for event-driven agents, which start the moment something happens, like a new job arriving or a document coming back from a court. The other is for scheduled agents, which run on a timer: hourly, daily, or weekly. Each agent lives in its own folder with a standard structure: a JSON config file, a system prompt in Markdown, deployment scripts, and operational documentation. Merging a change into the main branch deploys the agent automatically. A builder never has to write software. They clone the repo, copy a starter template, tell Claude Code what the agent should do, and get back everything the agent needs: config, prompt, credential store, and memory.

Bridging the technical divide

Fuller gathered the company’s 15-person steering committee, drawn from finance, marketing, operations, and development (none of them software developers), and had them clone the repository and build Managed Agents using Claude Code.

The goal was to prove that non-developers could build production agents themselves. If every agent had to route through the dev team, that bottleneck would cap how fast the whole company could move. What made it safe is that they were not writing software. Instead, they were filling in configuration and a prompt, and Managed Agents supplied the runtime.

"I had to explain what a PR was to them. A lot of [the non-software engineers] thought it meant running, like a PR, the fastest you can,” he said. “Now they're doing pull requests and sending them to each other."

Within a week, all 15 employees had working agents. Those builders went back to their teams and trained others. Within a month, roughly 50+ agents were running across ABC Legal. Each agent has a name, an owner, and a single job.

An agent for most stages of the legal document process

ABC Legal now has an agent at most stages of the legal filing process and the operations around it.

The AI Code Reviewer reviews every pull request across four codebases, running multi-model analysis to catch security bugs, performance regressions, and committed credentials. Engineers now wait for its review before merging.

The EvidenceChain™ Delivery Agent took over a weekly chore an account manager used to do by hand. ABC Legal runs a proprietary site, EvidenceChain.com, where courts, plaintiffs, and defendants look up the record of a service completed in the field, including who the process server was, when they attempted it, and photos of the document delivery. One customer wanted specific records pulled from it on an ongoing basis. The agent now pulls a database report for matching jobs, retrieves each PDF with a browser built into the Managed Agent, and delivers it to the customer's FTP server daily. The account manager who set it up had never automated anything, and built it in about an hour by describing it to Claude Code.

The eFiling Rejection Diagnoser fires automatically when a court rejects a filing, reads the job details, checks the court's rules, and posts a diagnosis to Slack in about a minute, work that used to consume hours of an employee’s day. A job-verification agent checks every incoming job against the courts. It navigates a court website in a browser, confirms the hearing or case is filed appropriately and actually occurring on the stated date, then adjusts the job based on what it found, flagging jurisdictions, courts, and statute-of-limitations timeframes.

The Attorney Coverage Agent works the network of attorneys to get hearings covered, checking availability, emailing them, and reading replies about availability and pricing so a coordinator can confirm coverage.

In finance, an AR-remittance agent parses a remittance email, builds the NetSuite payment-application file, and posts it to Slack for one-click approval, and then imports it, with a daily agent that renders a capitalize-or-expense verdict on each engineering ticket. Marketing runs a Google Ads analyst that posts a weekly recommendation for the channel lead. In operations, a review agent called Charvis checks completed service jobs and now agrees with the compliance team about 98% of the time.

The Service-Overdue-Nudger works the tier-1 layer of ABC Legal's operational backlogs, the repetitive first pass a person would otherwise do, and drafts tiered daily outreach messages for human approval.

Making the agents smarter: harvest, tune, repeat

ABC Legal's agents work under human supervision, posting what they did or what they recommend to Slack, where people reply in threads and react with emoji.

Hank, an internal code review agent, posts every review to a shared Slack channel. Each entry names the pull request and the counts that came out of it so the trail of what the agent decided is public and searchable.

Fuller saw all that reaction data as a training signal going to waste. Not every agent needs the signal, though. Most of the fleet are single-task runners whose output no one grades, and they work alone. For the agents that do collect graded feedback, ABC Legal uses a three-role architecture: separate agents that share one workspace, environment, and credential vault but run on different schedules. The pattern turns messages in Slack into versioned, human-approved changes to the agent:

1. The Initial Agent does the work, usually in real time as a job comes in or a document comes back, and records an audit trail of each action.
2. The Harvester runs hourly or daily and gathers human feedback from Slack, where it arrives as thread replies and emoji reactions. Each one becomes a labeled data point.
3. The Tuner runs weekly, looks across everything at once, and proposes a change to the prompt or config rather than the model's weights. It drafts only. A human reviews and merges the pull request.

In ABC Legal’s self-improving agent loop, an initial agent does the work in real time, a harvester sweeps up human feedback from Slack on an hourly cadence, and a weekly tuner proposes prompt and config changes as a pull request. Agents improve through the same workflows developers already use.

One example is "deliveries-as-code," Fuller's agentic system for tuning how work gets routed, which started at Docketly, ABC Legal's 50-person sister company. Docketly organizes its work around deliveries, each with its own ruleset for routing and handling. All 145 or so rulesets are single YAML files in git rather than records in an admin screen, so tuning a delivery means editing a file and opening a pull request.

Four agents make up the loop: one posts a weekly verdict to Slack, the Harvester turns reactions into labels based on human feedback, the Tuner opens a pull request on the YAML, and a fourth agent pushes the merged config to the production database. That fourth agent only executes what a human has already reviewed and approved. In practice, an emoji reaction flagging a mis-routed delivery can become a merged change to that delivery's routing rules within the week. The review is the only manual step in the loop.

Why Claude Managed Agents

Fuller evaluated multiple frameworks before settling on Claude Managed Agents as his organization’s agentic harness. His criteria were specific: the platform had to have versioning, observable sessions, workspace billing, model selection, memory primitives, MCP wiring, and, most critically, no infrastructure to babysit.

The platform's division of responsibility maps cleanly to how Fuller wants to run things. Anthropic’s managed infrastructure owns everything that makes an agent run: the execution loop, sessions, memory, the console, and the models themselves. ABC Legal owns the prompt, the tool list, the trigger logic, the audit trail, and the feedback loop on outcomes.

A few capabilities proved especially important at scale:

• Versioning: every push creates a new agent version with optimistic locking. Rollback is trivial.
• Model flexibility: the default is Claude Sonnet for most agents, Claude Haiku for high volume and fast tasks, and Claude Opus when deeper reasoning justifies the cost. Swapping models is a one-line change.
• MCP wiring and credential vaults: agents connect to ABC Legal's own platform (with over 100 tools available), Metabase for reporting, Slack for human-in-the-loop interaction, and Atlassian for project management.
• Scheduled deployments: recurring agents run on cron schedules through Bitbucket Pipelines, which already handles repo access, secrets, and billing.

ABC Legal tracks every dollar of AI spend, broken out by vendor, tool, team, and use case. Spend climbed as the fleet went live through the spring, then started falling in July while usage kept growing, the result of the efficiency work described below, with a ~50% reduction in cost for the tasks many agents cover and ~310 employees across every department using Claude.

The company's approach to cost is deliberate: push spend toward vertical, operational tools and agents where return is measurable, whilef keeping horizontal chat and ideation usage broad and costs in check. Most agents start with a human in the loop, where the agent looks at the job or ticket and makes a recommendation for a person to review before anything is acted on. The recommendation is either stored in the job and surfaced in a banner so the person can accept or reject it in the flow of their work, or posted to a Slack channel where people can reply in the thread. Those responses build a labeled dataset of good and bad calls, which feeds the harvester and tuner loop and lets the team write evals and benchmark agents across frontier models. Once an agent proves it is as good as or better than the humans on that specific task, it shifts into automation mode and acts on its own, and it stays inside the same measurement framework afterward to watch for any changes in performance.

The metric ABC Legal tracks is an efficiency ratio; the value an agent delivers measured against what it costs to run. Every Managed Agent reports its own value back to a data warehouse on each run, in hours and dollars. Agents follow a J-curve, often starting underwater while they are new and running larger models, then flipping positive as the team writes evals, moves to cheaper and faster models, and trims tokens.

Best practices for deploying a fleet of agents

Fuller’s experience with deploying AI–specifically Claude Managed Agents led him to a few working principles about using the technology:

‍

• Think of everything as code. "Code is just structured text. LLMs are text engines,” he said. “The more of your business you can turn into text in a repo, the more leverage agents give you." This applies to traditional software and equally to prompts, schemas, dispatch rules, notification templates, and business configurations.

• Start with humans in the loop. Every agent begins by posting recommendations for human review. Only after demonstrating consistent agreement with human decisions does it earn the right to act independently. "Every agent earns trust before it acts alone. It doesn't start there."

• Use the PR as your control surface. "If you want an agent involved in a decision, make the decision look like a pull request." Line-by-line comments, approval workflows, and immutable audit trails come free with version control, and compose naturally with both AI and human review.

• Invest in the feedback loop. The harvester-tuner pattern means agents improve without retraining. Slack replies and emoji reactions become structured signals that feed back into prompt and config changes, all through the same pull request workflow humans already use.

• Skip the scheduled-tasks detour. ABC Legal spent real time building scheduled tasks and local routines before moving to Managed Agents, largely because the product had only just launched in beta. Fuller's advice today is to go straight to Managed Agents.

• Expect the git hurdle, not the AI hurdle. The hard part was getting business users comfortable with cloning a repo and working in Git and pull requests, more than anything about the AI itself. It worked, and fast, but it was a real hurdle, and Fuller would like to see it made easier in the tooling itself.

• Not every task deserves an agent. The cost is real, so every team has to think in terms of value over cost. The work is picking tractable problems that genuinely save time or create automation, and being willing to say a given task is not worth an agent.

What's next

ABC Legal's agent fleet continues to grow. In-flight projects include a service photo reviewer, a PagerDuty triage agent, a daily KPI digest, and expanded Tuner loops on existing agents.

The team is also identifying more "X-as-code" candidates: notification templates, event routing rules, and dispatch logic that can be moved into repositories where agents can read, reason about, and propose improvements.

As Fuller puts it: "We want AI to support a business that can run itself, with employees free to steer it."

Learn more about Claude Managed Agents.

‍