Week 1: Building the First Agent (The Messy Reality)

I thought deploying the first Ruray agent would take a weekend. It took six days.

Day 1: I set up the development environment, installed the Anthropic SDK, wrote the initial intent classification prompt. It worked beautifully in testing. I was confident.

Day 2: I connected it to the real Instagram DM API. The first message it received was in Italian. My language detection returned 'uncertain.' It defaulted to English. The sender was confused. Back to the prompt.

Day 3–4: Edge cases. People who write in three languages in one message. Voice note transcriptions with typos. Someone who sent only an emoji. Each one required a rule, a fallback, a graceful recovery path.

Day 5: I introduced the gym manager to the shadow mode concept — the agent drafts, Felix approves before sending. She was skeptical. 'What if it says something wrong?' Valid fear.

Day 6: We reviewed the first 12 hours of shadow mode together. The agent handled 8 out of 10 messages correctly. The two it got wrong were both edge cases I hadn't seen before. We adjusted the prompts. We went live on Day 7.

This is what 'deploying in 14 days' actually looks like. Not a weekend. Six days of real work, real edge cases, and one very patient partner who trusted the process.