How ABC Legal Turned Every Employee into an AI Agent Builder with Claude
ABC Legal deployed Claude Managed Agents with a GitOps harness, enabling 15 non-engineers to build 50+ production agents in a month, featuring self-improving feedback loops via Slack, achieving 50% cost reduction on some tasks and 310 daily users.
Enterprise AI Agent Adoption at ABC Legal
ABC Legal, a U.S. legal document delivery company with 1,100 employees, adopted Claude Enterprise and later Claude Managed Agents to transform scattered AI experiments into a governed, professional agent workforce. CTO Brandon Fuller rolled out Claude Enterprise early in 2026; teams across litigation, e-filing, attorney operations, marketing, compliance, and finance immediately began automating tasks that previously consumed hours each day.
From Enthusiasm to Engineering: Treat Every Agent as Code
Fuller mandated that every agent be defined as code. He argued that an agent is merely structured text — a prompt plus configuration — and any text can live in a Git repository for visibility, review, and improvement. Each agent's prompt, tool list, schedule, credentials, and memory are stored in configuration files alongside the company's software. All changes require an approved pull request, giving every agent version history, code review, rollback, and audit trails.
Fuller spent a week building a starter kit with two templates in a dedicated Git repository: one for event-driven agents (triggered by new tasks or court documents) and one for scheduled agents (running hourly, daily, or weekly). Each agent lives in its own folder with a standard structure: a JSON config file, a Markdown system prompt, a deploy script, and runbook documentation. Merging to main triggers automatic deployment. Builders never write software; they clone the repo, copy a starter template, describe the agent's job to Claude Code, and receive the complete configuration, prompt, credential storage, and memory.
Bridging the Technical Gap
Fuller assembled a 15-person steering committee from finance, marketing, operations, and development — none were software engineers. They cloned the repo and used Claude Code to build managed agents. The goal was to prove non-developers could ship production-grade agents without an engineering bottleneck. Safety came from the fact that they were not writing code; they were filling configuration and prompts while Managed Agents provided the runtime.
Fuller had to explain what a pull request was; many thought "PR" meant a personal running record. Within a week, all 15 had working agents. These builders returned to their teams and trained others. Within a month, roughly 50 agents were running internally, each with a name, an owner, and a single responsibility.
Agents Covering Most Stages of the Legal Filing Workflow
AI Code Reviewer reviews every pull request across four repositories, running multi-model analysis to catch security vulnerabilities, performance regressions, and committed credentials. Engineers now wait for its approval before merging.
EvidenceChain™ Delivery Agent automates a weekly manual task for an account manager. It pulls database reports for matching jobs, uses the Managed Agent's built-in browser to retrieve each PDF, and sends them daily to the client's FTP server. The account manager, who had never automated anything, built it in about an hour by describing the need to Claude Code.
eFiling Rejection Diagnoser triggers when a court rejects a filing, reads job details, checks court rules, and posts a diagnosis to Slack in about a minute — work that previously consumed hours of staff time.
Job Validation Agent checks every incoming job against court information. It navigates court websites in the browser, confirms hearings or cases are properly filed and will occur on the stated date, then adjusts the job, tagging jurisdiction, court, and statute-of-limitation deadlines.
Attorney Coverage Agent manages the attorney network to ensure hearing coverage; it checks availability, emails attorneys, and reads replies about availability and quotes so coordinators can confirm coverage.
Finance agents : an AR Remittance Agent parses remittance emails, generates NetSuite payment application files, posts to Slack for one-click approval, then imports; a daily agent classifies every engineering ticket as capitalized or expensed.
Marketing runs a Google Ads Analysis Agent that posts weekly recommendations for channel leads.
Operations : Charvis, a review agent, checks completed serve jobs and now agrees with the compliance team's judgment about 98% of the time.
Service-Overdue-Nudger handles tier-one of the operations backlog — the repetitive first-pass work originally done by humans — and drafts tiered daily outreach messages for human approval.
Making Agents Smarter: Collect, Tune, Repeat
ABC Legal's agents operate under human supervision, posting completed work or recommendations to Slack where people reply in threads and react with emojis. Most agents are single-task runners with no scoring and no shared learning. For agents that do collect scored feedback, ABC Legal uses a three-role architecture: different agents share the same workspace, environment, and credential vault but run on different schedules. This pattern turns Slack messages into version-controlled, human-approved agent changes:
Initial Agent does the work, usually in real time as tasks arrive or documents return, and logs an audit trail for every action.
Harvester runs hourly or daily, collecting human feedback from Slack thread replies and emoji reactions. Each piece becomes a labeled data point.
Tuner runs weekly, looks at all the data holistically, and proposes changes to prompts or configuration — not model weights. It only drafts; a human reviews and merges the pull request.
An example is "deliveries-as-code," Fuller's agent system for tuning work routing. An emoji reaction flagging a delivery routing error can become a merged change to that routing rule within a week. The harvester turns reactions into labels; the tuner opens a PR against the YAML; a fourth agent pushes the merged config to the production database. The fourth agent only executes what has been human-reviewed and approved. Review is the only human step in the loop.
Why Claude Managed Agents
Fuller evaluated several frameworks before choosing Claude Managed Agents. His criteria were specific: the platform must provide versioning, observable sessions, workspace billing, model selection, memory primitives, MCP wiring, and — most critically — require no infrastructure babysitting.
The platform's division of responsibilities matched how Fuller wanted to operate. Anthropic's managed infrastructure handles everything needed to run agents: execution loops, sessions, memory, console, and the models themselves. ABC Legal owns the prompts, tool lists, trigger logic, audit trails, and the feedback loop.
Several capabilities proved especially important at scale:
Versioning: Every push creates a new agent version with optimistic locking. Rollback is trivial.
Model flexibility: Most agents default to Claude Sonnet; high-frequency, fast tasks use Claude Haiku; when deeper reasoning justifies the cost, they switch to Claude Opus. Swapping models is a one-line change.
MCP wiring and credential vault: Agents connect to ABC Legal's own platform (100+ available tools), Metabase for reporting, Slack for human-in-the-loop interaction, and Atlassian for project management.
Scheduled deployments: Periodic agents run on cron schedules via Bitbucket Pipelines, which already handles repo access, secrets, and billing.
ABC Legal tracks every dollar of AI spend, broken down by vendor, tool, team, and use case. Spend spiked when the agent fleet launched in spring, then began declining from July onward even as usage grew. This was driven by efficiency gains: many agents reduced the cost of covered tasks by roughly 50%, while about 310 employees across departments used Claude daily.
The company's cost approach is deliberate: push spend toward vertical operational tools and agents where ROI is measurable, while keeping horizontal chat and brainstorming usage broad but cost-controlled. Most agents start with human-in-the-loop: the agent reviews a task or ticket and posts a recommendation for human review before acting. Recommendations are either stored in the task with a banner for staff to accept/reject in their workflow, or posted to a Slack channel for threaded replies. These responses build a labeled dataset of good and bad judgments that feeds the harvester-tuner loop and lets the team write evals to benchmark agents across frontier models. Once an agent proves it matches or exceeds human performance on a specific task, it switches to autonomous mode but remains under the same measurement framework to monitor for drift.
ABC Legal tracks an efficiency ratio: value delivered by the agent divided by cost to run it. Every Managed Agent reports its own value back to the data warehouse per run, measured in hours and dollars. Agents follow a J-curve: they often start negative because they are new and run larger models; then, as the team writes evals, migrates to cheaper faster models, and trims tokens, they turn positive.
Best Practices for Deploying an Agent Fleet
Treat everything as code. "Code is just structured text. LLMs are text engines. 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, routing rules, notification templates, and business config.
Start with human-in-the-loop. Every agent begins by posting recommendations for human review. Only after proving it consistently aligns with human decisions does it earn the right to act independently. "Every agent has to earn trust before it acts on its own. It doesn't start that way."
Use PRs as your control plane. "If you want an agent to participate in a decision, make that decision a pull request." Line-by-line comments, approval workflows, and immutable audit trails come free with version control and work 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 PR workflow humans already use.
Skip the cron detour. ABC Legal spent months building cron jobs and local routines before moving to Managed Agents, largely because the product had just launched in beta. Fuller's advice now is to start directly with Managed Agents.
Expect a Git learning curve, not an AI learning curve. The real friction is getting business users comfortable cloning repos and working in Git and pull requests, not the AI itself. It works and it's fast, but it's a real hurdle Fuller hopes tooling will make easier.
Not every task deserves an agent. Cost is real, so every team must think in terms of value exceeding cost. The key is picking tractable problems that genuinely save time or enable automation, while also being willing to admit a particular task isn't worth an agent.
What's Next
ABC Legal's agent fleet continues to grow. Active projects include a Service Photo Reviewer, a PagerDuty Triage Agent, daily KPI summaries, and expanding the tuner loop across existing agents.
The team is also identifying more "X-as-code" candidates: notification templates, event routing rules, and distribution logic that can move into the repo so agents can read, reason about, and propose improvements to them.
As Fuller puts it: "We want AI to back a business that runs itself, so our people can steer."
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
