How Coupa freed engineers from manual work with production AI agents on AWS

Like many enterprise software companies, Coupa, an AWS Partner, ran a development pipeline that relied heavily on manual effort. Quality engineers would begin their day by manually triaging overnight test failures and logging tickets. When production issues occurred, engineers needed to piece together signals from separate logging, monitoring, and alerting tools to find the cause. This effort could take weeks.

No layer of shared intelligence connected the systems that held the answers—across continuous integration and delivery (CI/CD) pipelines, issue response, and customer security questionnaires. On their own, those tools were capable, but nothing could reason across them or act on what they found. Coupa saw a path forward in AWS.

First, the company needed leading foundation models and cost-effective infrastructure to run intelligent agents through a single interface inside the existing security perimeter. So it used Amazon Bedrock, a service for building generative AI applications and agents at production scale.

Putting agents into production was the harder part. Agents that open pull requests, file tickets, and query live databases require secure authentication, session persistence, and governance. And building that infrastructure in-house would have taken months. Amazon Bedrock AgentCore, a solution for building, connecting, and optimizing agents, delivered it out of the box.

Coupa built Sentinel, an internal solution that gives AI agents a single, coordinated way to operate across engineering tools. Sentinel can access all the models in AWS Bedrock through one API. Sentinel’s intelligent, task-based routing reserves the most capable models for complex reasoning and sends high-volume work to lighter ones.

The better grounded an agent is, the more autonomously it can act, so Coupa wanted to ground agents in context that’s specific to its engineering organization. Sentinel uses Amazon Bedrock Knowledge Bases to help applications and agents make better decisions based on organizational information.

“We’re processing 3–5 billion tokens a month,” says Bradley Simpson, director of engineering and head of the Sentinel platform at Coupa. “The ability to route intelligently across models on Amazon Bedrock is not a nice-to-have at that volume. It’s a meaningful cost and performance lever.”

Coupa prototyped Sentinel on a general-purpose framework years earlier, but as the solution scaled, the team moved its reasoning layer to Strands Agents. Taking a model-driven approach, the open source toolkit suits both Coupa’s multistep tool calling and its AWS based architecture. This framework determines which tools each agent calls, in what order, and when a task is complete.

Using Amazon Bedrock AgentCore, the team can run agents securely and autonomously against live production systems, in line with the AWS Shared Responsibility Model. Coupa draws on OAuth-governed access to internal tools, agent lifecycle management, session persistence that carries context across workflows and systems, and the audit trail necessary before an agent writes to production. Persistent memory keeps what an agent learns from one issue available to the next. The team doesn’t need to rebuild these controls every time it adds an agent because they live in the managed layer rather than inside agents.

“Amazon Bedrock AgentCore is what makes this a production solution rather than a collection of scripts,” says Simpson. “Our agents take real actions: opening pull requests, creating tickets, and querying databases. Today, about 100 percent of Sentinel’s agent invocations flow through Amazon Bedrock AgentCore as our primary production pathway.”

Using AWS, Coupa cut ticket investigation time by 50 percent and the mean time to resolve issues by 30–50 percent. In support channels, Sentinel responds to virtually 100 percent of incoming requests in under 1 minute and resolves about 75 percent of them autonomously. Responding to customer security and compliance questionnaires, which once consumed 120–150 hours of specialist time, is now 65 percent more productive per engagement.

Sentinel serves more than 500 users and runs 234 agents across more than 150 active workflows in production, processing over 15,600 automated tasks weekly. One of the production use cases, Adaptive Model Selection, has demonstrated 78 percent cost reduction per session.

The company is now bringing agents into critical parts of the lifecycle, working toward agentic CI/CD and autonomous issue remediation. “What Coupa has built with Sentinel is not a proof of concept or a pilot program,” says Simpson. “Agents act on production systems every day within the same security and governance controls that our engineers work under. That’s what makes it something the engineering organization can depend on, and it’s growing.”