Agents in the Wild

This directory is no longer updated. Last update: November 16, 2026.

Issue #5 · Nov 9, 2026

A page that disagrees with itself, and 4 more agent deployments

Hi,

This week we added 10 new sourced use cases. The directory now has 61: https://agentsinthewild.co

This week's 5

1. TRM Labs' security agent closed 17% of alerts without paging anyone

One Slack agent handles IT helpdesk, alert triage, infrastructure PR review and patching. TRM counted about 10,000 interactions in 30 days. It closed about 238 alerts alone.

The catch: TRM gives no false-close rate. The 17% says how many alerts closed, not whether each close was right.

→ https://agentsinthewild.co/use-cases/trm-labs-security-agent-four-workflows

2. Salesforce's SATA agent agreed with analysts about 95% of the time

SATA does first-line triage on several hundred security alerts a day. Low-confidence cases stay with people.

The catch: The 95% comes from replaying past cases before launch. It is agreement with analysts, not proven accuracy in production.

→ https://agentsinthewild.co/use-cases/salesforce-sata-security-alert-triage

3. EvenUp reports personal injury drafts in about 30 minutes, down from 8 to 15 hours

The agent reads records that often run past 1,000 pages and cites a source page for each fact. An attorney reads, checks and signs every draft.

The catch: The page disagrees with itself. The headline says 15 hours to 15 minutes. The body says 8 to 15 hours to about 30 minutes.

→ https://agentsinthewild.co/use-cases/evenup-claude-injury-case-drafting

4. GitHub's Qubot answers data questions in plain language

GitHub staff ask about the data warehouse in Slack, VS Code or the Copilot CLI. In GitHub's experiments, curated context made Qubot three times faster at reaching the right answer.

The catch: That compares context setups, not Qubot and a human. GitHub gives no accuracy rate.

→ https://agentsinthewild.co/use-cases/github-qubot-internal-data-agent

5. Qonto's agents prepare transfers, and the customer approves each one

Qonto reports bank transfers 2X faster and payroll 5X quicker. Customers have delegated transfers as large as €100,000.

The catch: Anthropic published this, with Qonto's own numbers. The page gives no method or error rate.

→ https://agentsinthewild.co/use-cases/qonto-claude-banking-agents

Also new: Spellbook (about 530,000 contract reviews a month), AWS's Kiro alarm triage, Airtable's data agent, and two recruiting screeners from Intellias and Peoplebank. See them all at https://agentsinthewild.co

Pattern of the week

The agent prepares. A person presses the button. TRM's PRs end with a person clicking Merge. Qonto's customers approve each transfer. EvenUp's attorneys sign each draft. Kiro's agent never resolves a ticket.

Prompt you can copy

From the Kiro entry: make the agent rule out other causes first. This prompt is ours, not theirs. Give the agent read-only credentials.

Alarm: {alarm}
Read-only tools: {tools}

List at least three possible causes.
Run queries to test each one. For each claim, show the query and result.
Say which causes you ruled out, and why.
Post a diagnosis and a confidence level. Do not resolve the ticket or change its severity.

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

Get the next issue

One short email a week.

The newsletter is paused. Past issues stay in the archive.