5 AI agents that actually shipped (and what the numbers hide)

Hi,

Welcome to the first issue of Agents in the Wild. Every week: real AI agent deployments, the source, and the catch. No predictions, no "10 ways AI will change everything".

The directory now has 35 sourced use cases: https://agentsinthewild.co

This week's 5

1. Stripe merges 1,000+ agent-written pull requests every week

Engineers start a task from Slack and get back a tested pull request. Humans review every one, but none contain human-written code.

The catch: PR counts say nothing about how big or hard each change is.

→ https://agentsinthewild.co/use-cases/stripe-minions-coding-agents

2. Klarna's AI did the work of 700 agents. Then Klarna hired humans again.

The assistant took two-thirds of support chats in month one. Fifteen months later, the CEO said a cost-first push had hurt quality.

The lesson: measure repeat contacts and complaints, not just deflection.

→ https://agentsinthewild.co/use-cases/klarna-ai-assistant-customer-service

3. Octopus Energy's AI emails beat human replies on satisfaction (76% vs 72%)

A careful trial: routine emails only, about 4% of volume, and vulnerable customers always go to a person.

Why it's useful: it's a fair comparison on the same type of question.

→ https://agentsinthewild.co/use-cases/octopus-energy-arlo-emails

4. Airbnb migrated 3,500 test files in 6 weeks instead of 1.5 years

Each file went through fixed steps with a retry loop that fed errors back to the model. Their lesson: many retries beat a perfect prompt.

The catch: going from 75% to 97% automated was the hard part.

→ https://agentsinthewild.co/use-cases/airbnb-enzyme-rtl-migration

5. Google's AI co-scientist matched a superbug hypothesis that took researchers years

Six agents propose, critique, and rank ideas. It reached the Imperial College team's unpublished answer in about two days.

The catch: one test, with a known answer. Promising, not proof.

→ https://agentsinthewild.co/use-cases/google-ai-co-scientist

Pattern of the week

The deployments that work have a check the agent can't fake: tests that pass, a numeric match, a customer who says "yes, solved". The ones that disappoint optimize a number that is easy to move (deflection, PR count) without a quality check next to it.

Prompt you can copy

From the Airbnb entry: a migration prompt meant to run inside a retry loop.

Convert this test file from Enzyme to React Testing Library.
Rules:
- Test behavior the user can see, not component internals.
- Use screen queries (getByRole first).
- Keep every existing test case. Do not delete assertions.

{file}

If retrying, the last attempt failed with this error. Fix only what the error requires:
{error}

Know a real deployment?

If you've seen (or built) an agent in production with a public write-up, send it in. Vendor case studies are welcome; we label them. → https://agentsinthewild.co/submit

Thanks for reading, Agents in the Wild