5 new AI agent deployments, and what each one doesn't prove

Hi,

This week we added 5 new sourced use cases. The directory now has 40: https://agentsinthewild.co

This week's 5

1. eSentire's Claude agents ran 120,000+ security investigations in a year

Each case uses about 44 tool calls across endpoint, identity, network and cloud data. Senior analysts review critical findings and check a sample of verdicts. The company reports above 90% agreement with top experts.

The catch: "Agreement" is eSentire's own measure on a sample it chose. No outside audit.

→ https://agentsinthewild.co/use-cases/esentire-claude-soc-investigations

2. Nubank's card-delivery agent gained 37 points of AI NPS in A/B tests

Nubank's KDD '26 paper covers five support agents and the test pipeline behind them. The same test shows a 29-point gain in self-service rate.

The catch: The gains are against Nubank's earlier agent versions, for one agent. Not against humans.

→ https://agentsinthewild.co/use-cases/nubank-support-agents-evaluation-framework

3. OpenRouter's sales agent returns about 600 hours a month to five people

It qualifies leads, drafts first emails, writes call briefs and fills the CRM. It costs about $30 a day, the company says. People keep the calls and the final say.

The catch: OpenRouter itself says pricing and market changes moved at the same time, so not all gains are the agent.

→ https://agentsinthewild.co/use-cases/openrouter-rasp-sales-agent

4. Coupa runs 234 engineering agents on one shared platform

The agents open pull requests, create tickets and query databases. AWS reports 75% of support queries resolved without a person.

The catch: It is an AWS customer story. Baselines are not shown, and some figures differ inside the page.

→ https://agentsinthewild.co/use-cases/coupa-sentinel-engineering-agents

5. Monte Carlo had every team build an agent in one week: 150+ logged

One data scientist let Claude tune a prompt against 10 test scenarios. Accuracy went from 78% to 97%.

The catch: It is a small internal test set, and most projects have no hour figures.

→ https://agentsinthewild.co/use-cases/monte-carlo-ai-sprint-150-agents

Pattern of the week

Last issue's pattern was a check the agent can't fake. This week, three of five teams built the check first: eSentire's expert review, Nubank's judge-plus-A/B pipeline, Monte Carlo's eval set. The agent came second.

Prompt you can copy

From the Monte Carlo entry: tune a prompt against a small test set. This prompt is ours, not theirs.

Here is an agent prompt: {current_prompt}
Here are test cases with correct answers: {eval_cases}

Run each case, report accuracy and recall, and list the cases it got wrong.
Propose one small change to the prompt. Re-run and compare.
Stop when accuracy stops improving for 3 tries. Show every change you tried.

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