AI agent systems, built and run with checks

Case studies

What an AI agent is, in plain terms

An AI agent is a language model that can take steps, not just answer a question: read files, call a tool, write code, send a request, check the result and try again. That makes agents useful for work with many small steps. It also makes them risky. An agent can edit something nobody asked it to touch, report "done" when nothing was checked, or keep spending while nobody is watching.

So the part of an agent system we care about most is not the model. It is everything around it: what each agent is allowed to see and change, how its work is checked, what it may spend, and who approves the risky steps.

Project Swarm: our own multi-agent system

Project Swarm is the build system behind every Verrow project. It is open source under the Apache 2.0 license, created and maintained by Verrow, and hosted under our partner RDW Labz at github.com/RDW-Labz/project-swarm (opens in a new tab). How it works, in short:

• One coordinator, many workers. Any capable agent can coordinate. The coordinator splits a goal into scoped jobs, each with one deliverable, the exact files the worker may read and write, and a check that shows the job worked.
• Lanes by cost. Jobs route to cheap, mid or expensive tiers. A job set to auto moves up a lane after two failed or timed-out attempts, so the expensive models only see the problems that earned them.
• Many agents and providers. The runner has thirteen adapters, including Claude Code, Codex CLI, Cursor CLI, Hermes, Qwen Code, OpenRouter and any OpenAI-compatible endpoint.
• Review before merge. The coordinator inspects every result. A merge guard refuses any change unless every check is green, and there is no flag to skip it.
• A security gate. Written files are scanned for secrets, new dependencies must exist, be at least 30 days old and not imitate a popular name, and risky changes are held for the owner.
• Hard spend caps. A job starts only if it fits under the cap, and raising a cap needs the owner's approval.
• Every mistake becomes a test. Problems from real runs are written down as lessons and, where possible, turned into a check and a regression test.

We wrote up the details in How Project Swarm builds software with a team of AI agents.

Agent Exchange: agents as customers

Agent Exchange is a business we are building that is run by agents, for agents: small tools that other AI agents call over the network, priced per call, connected over MCP (the Model Context Protocol). It is in build and not live. The Agents page shows a simulation with example data and pricing, labelled as such.

AI agents for a small business

The same rules carry over to agent systems we build for a business:

• Scope. Each agent gets one job and only the data and tools that job needs.
• Checks. Every result is checked before it changes anything that matters, by a test, a rule or a person.
• Limits. Spend caps and rate limits are set before the first run, not after the first bill.
• A person on the judgment calls. Agents do the repetitive steps and hand the decisions to a person.
• Secrets stay out. Workers never see passwords or API keys.

Often the honest answer is that a business does not need an agent at all, and a scheduled job or a well-built app does the work more cheaply and more reliably. See AI automation and custom software for that side of our work.

How an agent project runs

1. Understand the work where it happens. What the job is, who does it now, and what a mistake would cost.
2. Build the first version you can use. One agent, on one job, with its checks, before anything else.
3. Build with our own agent system. The agent itself is built with Project Swarm, where every change is checked and tested before it merges.
4. Hand over, support and keep improving. You get the code and the notes, and we stay on to fix and improve it.

Questions people ask

How much does it cost to build an AI agent?

Two costs: building it, and running it. Running cost depends on the model and how often the agent is called, and it should be capped from day one. We do not publish a price list; we scope the job with you first and start with one agent on one job.

Which AI models do you use?

It depends on the job. Project Swarm can drive Claude, Codex, Cursor, Hermes, Qwen, Gemini and OpenAI models, models through OpenRouter, and models you host yourself. Before any model sees your code or data, we agree which providers may receive it.

Can I use Project Swarm myself?

Yes. It is free and open source. The repository (opens in a new tab) has a kickoff prompt to paste into your agent, and the Project Swarm page explains how a run works. If you want it set up on your project, email or call.

Is Agent Exchange live?

Not yet. It is in build. Email or call to get on the list.

Do you work outside Arkansas?

Yes. We are in Harrison, Arkansas, and work with businesses anywhere in the US by email and phone.

Other services

• Custom software: Operations apps, employee portals and compliance record apps, built around how your crews actually work.
• AI automation: Scheduled texts, records that fill themselves in, approvals and reports, with AI only where it earns its place.
• Websites: Fast marketing sites, from menu-first local pages to 3D and scroll-driven sites.