I Built an Agent That Does My Job for Me

I built an agent for Origami (YC F24) that does my job for me.

About a month ago, I was spending 20 hours a week prepping for customer calls. Digging through transcripts, cross-referencing data across Stripe, Posthog, Clerk, Hubspot, Supabase, and Vitally. It was the kind of work that feels productive but is really just manual labor disguised as diligence.

By early May, I cut my prep time to 20 minutes a day.

How?

I unified our fragmented customer data and built an agent that delivers everything I need in seconds.

Instead of manually copy-pasting transcripts into ChatGPT and jumping between six different dashboards, I created one system that instantly surfaces relevant information when you need it.

The Questions That Used to Take 30 Minutes

Think about how often these come up in any customer-facing role:

• "I have about 5 hours of calls with this client. I know they mentioned sensitivity to pricing once. What was it again?"
• "Where is this client based again?"
• "Prepare me for my call in 30 minutes. How have they been using the platform?"

Each of these used to require opening multiple tabs, scrubbing through call recordings, and piecing together context from memory. Now it's a single query.

Why This Matters Beyond Call Prep

As Origami scales, these internal agents will save us from hiring for positions we don't actually need. The kind of work that traditionally requires a dedicated ops person — data aggregation, context synthesis, prep work — can be handled by agents that never sleep and never forget.

This isn't about replacing people. It's about freeing people to do the work that actually requires human judgment.

The Unexpected Bonus

I've added voice features so I can get briefings on my bike ride to work. What used to be 45 minutes at a desk is now a conversation during my commute.

The best tools don't just save time — they change the shape of your day.

The Ripple Effect

This agent wasn't an isolated project. It was part of a progression where each build stacked on the last:

1. TIS (The Intelligent Search) — A research agent with MCP connectivity. Deployed on Render with 3 sync API routes across separate repos. This taught me how to structure multi-service projects.
2. Cold Outbound Consultant — Managed 25 client accounts with agents handling their outbound. Clients averaged 15-20% response rates.
3. Money-Collector — When I joined, we were collecting $5k/month. After building a system to track billing and automate collections, we hit $65k in a single month. That system still runs.
4. The Nest — The culmination. An MCP server with an open-source front end that automated call prep, invoice sending, billing tracking, and account management. Essentially automated my entire ops role.

Each project taught me something the previous one didn't. TIS taught me deployment. The outbound system taught me scale. Money-Collector taught me that automation compounds. The Nest taught me that the best engineering work eliminates your own job.

What I Learned Building It

1. Start with your own pain. I didn't build this because someone told me to. I built it because I was drowning in prep work.
2. Unify before you automate. The agent is only as good as the data it can access. Getting everything into one place was 80% of the work.
3. Ship ugly, iterate fast. The first version was a CLI tool that only I could use. That was fine. It proved the concept.
4. Push to Git more. Small, frequent commits. I learned this the hard way — going too long between pushes cost me hours of lost work more than once.
5. Testing isn't optional. As these agents handled real money and real client data, I realized that "it works on my machine" isn't good enough. Building testing habits early is the unlock I wish I'd prioritized sooner.

If you're an engineer at a startup, look at where your team spends the most time on repetitive knowledge work. That's where your first agent should live.