The AI Agent That Taught Itself 45 Tools and Helped Build a $2M Business With 2 Employees

The Setup

Two full-time employees. Two or three contractors for design and outbound. $2 million in annual recurring revenue.

That ratio , the one that would have been impossible to explain five years ago , is now the operating reality at Answer This, a startup building AI agents for scientific workflows. The founder presented the internal AI ops architecture at a YC Root Access event. Here is how it actually works.

What the Agent Handles

100-plus emails per day processed. 400-plus customer support tickets closed. CRM updates after every meeting. User feedback collected across channels. All of this runs through a single main agent that the team queries with natural language.

The agent has read-only access to both the database and the codebase. A cron job updates this copy every time there is a new release. When a customer asks a question that requires understanding the product's subscription logic or feature structure, the agent reads the codebase directly to answer it , accurately, without guessing.

The Part That Is Different From Every Other Agent Setup

The main agent has a CLI-based tool for general coding tasks. That coding tool can edit the main agent's own code.

When you ask the agent to do something it cannot do , a task that requires a tool or integration it does not have , it writes the tool itself and then uses it. The team started with a skeleton agent and gave it no predefined capabilities beyond the basic ones. Over time, through normal operation, the agent built 45 CLIs for itself.

One example: the team mentioned they wanted to monitor landing pages for uptime during ad campaigns. They told the agent. It created a cron job that monitors the pages and alerts the team when something goes down. Nobody wrote that cron job. Nobody specified how it should work. The agent figured out what was needed and built it.

The Personality Layer

Every turn, the agent loads an `instructions.md` file. This file contains the agent's personality and behavioral guidelines , how it should communicate, what tone to use, what it knows about the business and the team.

The agent can edit this file. When the team wants to change how the agent behaves, they can either edit the file directly or tell the agent what to change and let it update itself.

This is not a quirk of implementation , it is the design. The founders wanted an agent that could be steered without requiring a developer to touch code. Any team member can modify the agent's behavior by editing a markdown file or having a conversation about it.

Why Claude Code as the Harness

The choice of Claude Code as the underlying execution environment is intentional. It already knows how to inspect files, run commands, and use CLIs from training. You do not need to teach it how to navigate a filesystem or execute a shell command , that knowledge is baked in.

The founders describe it as a "thin harness." All the integration complexity lives in the tools the agent has built for itself. The foundation is deliberately minimal.

The architecture is replicable. None of it requires proprietary infrastructure or specialized ML expertise. What it requires is a willingness to give the agent real tools, real access, and enough time to encounter tasks it cannot yet handle , because those are the moments when it grows.