Agents in the Wild

This directory is no longer updated.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

TRM Labs' security agent handled about 10,000 interactions in 30 days and closed 17% of alerts alone

TRM Labs runs one Slack-based Claude agent across IT helpdesk, security alert triage, infrastructure PR review and vulnerability patching. People still click Merge on every code change and confirm every irreversible action.

Organization
TRM Labs
Industry
Blockchain intelligence
When
Jun 2026
Evidence
Company reported

Reported outcome

In the past 30 days: about 10,000 interactions, 45% of helpdesk requests resolved with no ticket, and about 238 security alerts (17%) closed without paging anyone.

  • ~10,000
    Interactions, past 30 days
  • 45%
    Helpdesk resolved with no ticket
  • ~238 (17%)
    Alerts closed without paging
  • USD 0.29
    Cost per alert triage

Caveats

Stated by the company that uses the agent (blog, earnings call, interview). Not independently audited.

  • TRM reports these numbers about its own system. Nobody outside TRM has checked them.
  • The post gives no false-close rate, so 17% says how many alerts closed, not whether each close was right.
  • It covers one 30-day window. Rates vary by workflow, and TRM gives no single platform-wide figure.
  • Agent PR approval covers only a small set of low-risk patterns. Every code-changing PR still ends with a person clicking Merge.

What the agent does

TRM's security team runs one agent in Slack with four jobs: IT helpdesk, security alert triage, infrastructure pull request review, and vulnerability patching.

Claude Opus acts as the dispatcher in a tool-use loop. The system exposes 56 tools across 14 skills, and each channel sees only a subset. Patch work runs as separate Claude Code workers that open pull requests.

What happened

Over the past 30 days, TRM counted about 10,000 interactions. The helpdesk reached 45% resolution in the strict sense, meaning no ticket was created. Alert triage closed about 238 alerts (17%) without paging anyone and escalated the rest with added context.

TRM reports cost per interaction by workflow: USD 0.16 for helpdesk, USD 0.25 for PR review and USD 0.29 for alert triage.

What to copy

  • Give each channel a small set of tools. TRM found that a model choosing from all 56 tools decided worse.
  • Put irreversible actions behind a button a person clicks, outside the model's control.
  • Report the strict number. TRM counts helpdesk resolution only when no ticket exists.
  • Track cost per workflow, not one blended figure.

Starter prompt

Written by us, not by TRM Labs. Test it on past alerts before you let it close anything.

Alert: {alert}
Context from logs and asset inventory: {context}
Known benign patterns: {benign_patterns}

Decide: CLOSE or ESCALATE.
Choose CLOSE only if the alert matches a listed benign pattern and the context confirms it.
Quote the evidence for your choice.
If you are unsure, or a source is missing, choose ESCALATE and list what you checked.

Sources

  1. Scaling Security in the Age of AI, Part 2: How One Agent Triages, Remediates, and Approves Across TRM — TRM Labs, Jun 10, 2026

Added Nov 9, 2026 · Sources last checked Nov 16, 2026.

Get new use cases weekly

One short email. Sources included.

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

Related use cases