How ABC Legal Runs 50+ Claude Agents as Code (Case Study)

Fifty-plus Claude agents, defined entirely as text files in a git repository, reviewed by pull request like any other change to production software. That's the operating model Anthropic detailed on August 17, 2026 in a new customer story about ABC Legal, a U.S. legal document delivery company that runs Claude Managed Agents across service of process, eFiling, appearance counsel operations, marketing, compliance, and finance. @ClaudeDevs shared the story on X the same day, pulling in nearly 193K views within hours.

The headline numbers are the hook — 50+ agents, ~310 employees using Claude daily, and up to a ~50% cost reduction on the human tasks some of those agents cover. But the part worth stealing for your own org isn't the count. It's how ABC Legal's CTO, Brandon Fuller, made a 1,100-person company with a 15-person, mostly non-technical steering committee ship production agents in a week — and then kept them improving without anyone retraining a model.

TL;DR

The pattern: treat every agent like a piece of software

Fuller's starting premise is simple and, once you sit with it, kind of obvious: "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."

So that's what ABC Legal did. Every agent's prompt, tool list, schedule, credentials, and memory live in configuration files inside a git repository next to the company's regular software. Nothing about an agent changes except through a pull request someone approves — which gets every agent version history, code review, rollback, and an audit trail for free, the same guarantees CI/CD already gives conventional software.

Fuller spent a week building two starter templates, each living in its own git repository:

• Event-driven agents — start the moment something happens: a new job arriving, a document coming back from a court.
• Scheduled agents — run on a timer: hourly, daily, weekly.

Each agent's folder follows a standard shape: a JSON config file, a system prompt in Markdown, deployment scripts, and operational docs. Merging a change into the main branch deploys the agent automatically. A builder never writes software — they clone the repo, copy a template, describe the agent to Claude Code, and get back config, prompt, credential store, and memory, ready to deploy.

Proving non-developers could ship agents without a dev-team bottleneck

If every agent had to route through the engineering team, that team becomes the ceiling on how fast the whole company can move. Fuller tested the opposite bet directly: he gathered ABC Legal's 15-person steering committee — drawn from finance, marketing, operations, and development, none of them software engineers — had them clone the repo, and had them build Managed Agents using Claude Code.

It worked because they weren't writing software. They were filling in configuration and a prompt; Claude Managed Agents supplied the runtime underneath. Fuller's own framing of the learning curve: "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. Now they're doing pull requests and sending them to each other."

All 15 had working agents within a week. Those builders trained others on their teams, and within a month roughly 50+ agents were running across ABC Legal — each with a name, an owner, and a single job.

What the agents actually do

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

The self-improving loop: harvester, tuner, and Slack as a training signal

Most of ABC Legal's fleet are single-task runners whose output no one grades. But for agents where feedback is worth collecting, ABC Legal built a reusable three-role architecture — separate agents sharing one workspace, environment, and credential vault but running on different schedules:

1. The Initial Agent does the work in real time — a job coming in, a document coming back — and logs an audit trail of each action.
2. The Harvester runs hourly or daily and sweeps human feedback out of 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, and proposes a change to the prompt or config — never the model's weights — as a pull request. A human reviews and merges it.

This is the part of the story that generalizes past legal-ops: the harvester/tuner pattern turns reactions people are already giving in Slack into a versioned, human-approved training loop, without touching a model's weights or standing up an eval pipeline from scratch. It's a lightweight alternative to formal RLHF-style tuning that any team already running agents on Claude Enterprise or a similar agent harness could copy directly.

ABC Legal's clearest example is "deliveries-as-code," a system that started at Docketly, ABC Legal's 50-person sister company. Docketly organizes its work around deliveries, each with its own routing and handling ruleset. All 145 or so rulesets are single YAML files in git rather than rows in an admin screen, so tuning a delivery means editing a file and opening a pull request. Four agents make up that loop: one posts a weekly verdict to Slack, the Harvester turns reactions into labels, the Tuner opens a pull request on the YAML, and a fourth agent pushes the merged config to the production database — executing only what a human already reviewed. In practice, an emoji reaction flagging a mis-routed delivery can become a merged routing-rule change within the week, with human review as the only manual step.

Why Claude Managed Agents specifically

Fuller evaluated multiple frameworks before settling on Claude Managed Agents. His criteria: versioning, observable sessions, workspace billing, model selection, memory primitives, MCP wiring, and — most critically — no infrastructure to babysit.

The platform's division of responsibility is the part he called out specifically: Anthropic's managed infrastructure owns the execution loop, sessions, memory, the console, and the models. ABC Legal owns the prompt, the tool list, the trigger logic, the audit trail, and the feedback loop on outcomes. A few capabilities mattered most at scale:

• Versioning — every push creates a new agent version with optimistic locking, making rollback trivial
• Model flexibility — Claude Sonnet by default, Claude Haiku for high-volume/fast tasks, 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 (100+ internal tools), Metabase for reporting, Slack for human-in-the-loop review, and Atlassian for project tracking
• Scheduled deployments — recurring agents run on cron schedules through Bitbucket Pipelines, which already handled repo access, secrets, and billing

ABC Legal tracks every dollar of AI spend by vendor, tool, team, and use case. Spend climbed through the spring as the fleet went live, then started falling in July even as usage kept growing — the payoff of the efficiency work described above. The company's stated approach: push spend toward vertical, operational agents where return is measurable, while keeping horizontal chat and ideation usage broad but cost-controlled — the same operating discipline explainx.ai has covered in other Claude token cost-control case studies.

The metric ABC Legal actually watches is an efficiency ratio — value delivered against cost to run — reported per agent run, in hours and dollars, to a data warehouse. Agents tend to follow a J-curve: underwater at first while new and running larger models, then flipping positive as the team writes evals, moves to cheaper/faster models, and trims tokens.

Best practices Fuller shared

• Think of everything as code. "The more of your business you can turn into text in a repo, the more leverage agents give you" — applies to prompts, schemas, dispatch rules, notification templates, and business configuration, not just traditional software.
• Start with humans in the loop. Every agent begins by posting recommendations for human review; it earns the right to act independently only after demonstrating consistent agreement with human decisions.
• Use the pull request as your control surface. 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 through the same PR workflow people already use.
• Skip the scheduled-tasks detour. ABC Legal spent real time on local scheduled tasks before Managed Agents existed 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 cloning a repo and working in git and pull requests — not anything about the AI itself.
• Not every task deserves an agent. Cost is real; the discipline is picking tractable problems that genuinely save time, and being willing to say a task isn't worth automating.

What's next

In-flight projects at ABC Legal include a service-photo reviewer, a PagerDuty triage agent, a daily KPI digest, and expanded Tuner loops on existing agents. The team is also hunting more "X-as-code" candidates — notification templates, event-routing rules, dispatch logic — anything that can move into a repository where agents can read, reason about, and propose improvements.

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

Related on explainx.ai

• Claude Managed Agents: Dreaming, multiagent orchestration, and outcomes loops
• Code with Claude Tokyo: scheduled agents, vaults, dynamic workflows GA
• Claude Enterprise-managed authorization for MCP connectors
• What is an agent harness?
• What is MCP? Model Context Protocol guide
• Slash Fintech cut an $80K AI bill with Claude token cost control
• DoorDash Flux: cloud agents platform, 130k tasks a month
• Top 15 Claude Tag use cases for Slack teams

Primary source: Anthropic — How ABC Legal turned every employee into a builder with Claude Managed Agents (August 17, 2026) · @ClaudeDevs on X

Figures and quotes are ABC Legal's own, as published in Anthropic's August 17, 2026 customer story — self-reported, not independently audited. Learn more about Claude Managed Agents directly from Anthropic.