Agents in the Wild

Coupa's Sentinel runs 234 engineering agents; 75% of support queries resolved without a person

Coupa built Sentinel, an internal platform of AI agents on AWS that open pull requests, create tickets and query databases for engineers. AWS reports halved investigation time and 75% autonomous resolution.

Organization
Coupa
Industry
Business spend management software
When
2026
Evidence
Vendor case study
Agents and tools
Amazon BedrockAmazon Bedrock AgentCoreBedrock Knowledge BasesStrands Agents

Reported outcome

500+ users, 234 agents and 15,600+ automated tasks a week; 50% less ticket investigation time and 75% of support queries resolved autonomously.

  • 234 (company/AWS)
    Agents
  • 15,600+ (company/AWS)
    Automated tasks per week
  • -50% (company/AWS)
    Ticket investigation time
  • 75% (company/AWS)
    Support queries resolved autonomously

Caveats

Published by the company that sells the agent. Treat numbers as marketing until confirmed elsewhere.

  • Results are self-reported in an AWS customer story. The page does not describe baselines or how time was measured.
  • Some figures differ between the summary and the body text (for example, issue resolution time falls '50%' in one place and '30-50%' in another).
  • The 78% cost cut applies to one production use case, not the whole platform.
  • Agentic CI/CD and autonomous issue fixing are described as ongoing work, not finished results.

What the agent does

Sentinel gives engineers one place to run agents across their tools. The agents can open pull requests, create tickets and query databases. They are grounded in Coupa's own engineering documents. Access is governed with OAuth, session records and audit trails.

Why this works

Instead of one agent per team, Coupa built one shared platform with shared permissions and logs. Simple, high-volume work goes to lighter models. Hard reasoning goes to stronger ones.

What to copy

  • Put tool access, sessions and audit logs in one shared layer before you add many agents.
  • Route easy tasks to cheaper models.
  • Count resolved-without-a-person, not only answered.

Starter prompt

Written by us, not by Coupa or AWS.

You are an engineering support agent. Ticket: {ticket_text}
Search the knowledge base and recent incidents for matches.

Return:
1. Likely cause, with the document or incident you used
2. Confidence: high, medium, low
3. Suggested fix, or "ESCALATE" if confidence is low
Never change anything. Only propose.

Sources

  1. Coupa builds an agentic engineering platform with Amazon Bedrock — AWS, 2026

Added Oct 19, 2026. Spotted an error? Send a correction.

Get new use cases weekly

One short email. Sources included.

Related use cases