AI agents for business, built to do real jobs

An AI agent is software that does work on its own inside your business. Not a chat window someone asks questions. Software that takes a job, reaches into your systems, does it, and reports back.

They are good enough now to be given real work. Most businesses that try them still end up with an expensive demo, and it is almost never the model's fault.

Why most agent projects stall

A company decides to "add agents." Someone wires a model to a few tools, it does something impressive in a meeting, and six weeks later nobody trusts it with anything that matters.

The reason is that the agent was given capability without employment. A person you hire gets an identity, access appropriate to the role, a manager who sets limits, a budget, and a record of what they did. An agent usually gets none of that. It wears the login of whoever set it up, it either asks permission for everything or for nothing, and nobody can say what it costs or what it did last Tuesday.

So it never gets trusted with anything consequential, so it never does anything consequential, so the project gets cancelled.

What I build

Agents that do a specific job in your business, plus the five things that let you trust them with it.

1. Identity. Every agent is someone: a name, a role, an accountable owner, a signature on its work.
2. Authority. What it may do alone, what needs approval, and what is forbidden. Written down, not assumed.
3. Direction. Work arrives in a form it can execute, in priority order, and it can reach a person wherever they are when it needs judgment.
4. Accounting. What each unit of agent work costs, and how often it fails, so you can price the work and see what makes money.
5. Record. What was done, by whom, under what authority, at what cost. In a company that runs on agents, this history becomes the asset.

The specific agents depend on your business. Common starting points are sales development and outreach, client follow up and reporting, intake and triage, content production, research, and the back office work that eats a team's week.

Agents I have already built in public

You should not have to take an agent builder's word for it. A lot of mine are open source.

• Harvey, an autonomous AI sales agent. It finds prospects, writes cold emails, sends campaigns, handles replies and books meetings. Over 2,000 developers have starred it on GitHub.
• JARVIS, a voice-first assistant for your Mac. You talk to your computer, it talks back and gets things done.
• Codex on Crack, which lets one conversation plan, delegate to other models, and review their work. It is how I run long builds without burning through usage limits.
• Billr, invoicing with a built-in MCP server, so agents can do the billing work with you. Hosted, or self-host the open source version.
• Agent Ledger, a local workspace for understanding what your coding agents did and what it cost.

And two businesses that run on agents: INKRALLY, where GPT-6 runs the whole business, and Plastix, a marketing agency being rebuilt to run on agents.

How a build works

1. Pick the job. One process, chosen because it is valuable and ready, not because it makes a good demo.
2. Write it down. What the work is, what good looks like, the exceptions, and what must never happen. This document is what the agent works from.
3. Make the systems reachable. The agent gets the same access a person in that role has, inside the same boundaries.
4. Run it with a person watching. Every output is checked by someone who knows the work, and the agent never grades its own homework.
5. Hand it the loop. Once it is reliable, it runs on its own, with a defined trigger, output and moment where it hands back to a person.
6. Hand it to you. Your team learns to run it, change it and build the next one.

If you have a job in mind, tell me about it. If you are earlier than that and want to know where agents fit in your company at all, start with AI consulting.