An agent said 7.5x. A human said 1.75x. Plus 4 more deployments

Hi,

This week we added 6 new sourced use cases. The directory now has 51: https://agentsinthewild.co

This week's 5

1. Shopify's River agent coauthored about one in eight merged pull requests

Employees mention River in public Slack channels. It reads code, runs tests, opens pull requests and checks production traces. Shopify reports 59,918 sessions and 3,536 merged pull requests in one 30-day window.

The catch: Shopify counts sessions and merges, not code quality. The numbers come from a table River writes about itself.

→ https://agentsinthewild.co/use-cases/shopify-river-slack-agent

2. Advocate Health: AI-drafted prior authorizations, from up to 25 minutes to about 5

AI inside Epic drafts answers to payer questionnaires from the patient record. Pharmacy staff review and submit. As reported, requests reach payers about a day sooner.

The catch: We could not open the original Epic case study, only a news summary. "Up to 25 minutes" is the high end, not an average.

→ https://agentsinthewild.co/use-cases/advocate-health-epic-prior-auth-drafts

3. Carta Healthcare cut registry abstraction time by up to 66% with Claude

Claude drafts answers to clinical registry questions with citations. Trained abstractors check each one. Carta reports 98-99% agreement between reviewers.

The catch: The per-case example (30 to 15-22 minutes) is a 27-50% cut, smaller than the 66% headline. Both sources are Anthropic's.

→ https://agentsinthewild.co/use-cases/carta-healthcare-lighthouse-registry-abstraction

4. Works Human Intelligence cut an HR browser agent's cost per process by up to 97%

The agent drafts a change in the HR system and waits for the user to approve. Prompt caching, shorter prompts and a smaller model did most of the saving.

The catch: It is an AWS blog post about AWS tools, and it reports cost only. No accuracy numbers.

→ https://agentsinthewild.co/use-cases/works-human-intelligence-hr-agents-agentcore

5. FutureHouse's Robin proposed ripasudil as a dry AMD drug candidate

Agents read the literature, ranked hypotheses, proposed drugs and analyzed lab data. Humans ran the experiments.

The catch: On the same data, the agent's analysis said 7.5-fold. A human's analysis said 1.75-fold. And this is a cell test, not a clinical result.

→ https://agentsinthewild.co/use-cases/futurehouse-robin-dry-amd-ripasudil

Also new: Freshfields gave Claude to 5,700 staff (usage, not results yet): https://agentsinthewild.co/use-cases/freshfields-claude-firmwide-legal

Pattern of the week

The person checks the number that matters. Advocate staff submit the forms. Carta's abstractors make the final call. WHI's agent waits for approval. And Robin shows why: the agent and the human read the same lab data and got very different answers.

Prompt you can copy

From the Carta Healthcare entry: answer from records, with a time window. This prompt is ours, not theirs.

Question: {question}
Time window: from {start} to {end}
Documents (with dates): {documents}

Answer using only documents inside the time window.
Give: answer, the exact quote, the document name and date, and one sentence of reasoning.
If nothing answers the question, say "No evidence found".

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