TCS says its AgentCore test agents cut scenario effort by 88%

Tata Consultancy Services (TCS) says an agent platform it built on Amazon Bedrock AgentCore cut test-design and execution effort by 88% on a representative banking workload. The figures come from a case study AWS published on its machine learning blog on 10 November 2026, co-written with Varun Valsan, a generative AI team lead at TCS.

TCS runs testing for banking, financial services and insurance (BFSI) clients. The post says the slow part of a release is the chain after the code: reading requirements, designing scenarios, writing and maintaining automation, diagnosing failures and assembling evidence.

The number and how it was measured

According to the post, TCS compared its old process with the new platform on about 100 test scenarios. Manually, each scenario took about two hours of engineer effort to design, script and execute, about 200 hours in total. With the orchestrator, the average was about 14 minutes per scenario, about 24 hours in total. The post calls this a controlled internal benchmark and says results vary by workload, application complexity and team.

How it is built

Testers feed in Jira stories, business requirement documents, Swagger specifications, GitLab source folders or captured UI flows. The platform produces test scenarios, Gherkin feature files and executable Java or Python tests, runs them on Amazon ECS and reports per tenant.

• Specialised agents, built with the Strands Agents SDK and hosted in AgentCore runtime, each own one stage: generation, conversational refinement, requirement-to-story conversion, incremental generation across linked Jira stories, and execution.
• A deterministic supervisor routes work between them, rather than a model deciding the order.
• Project documents are indexed in Amazon Bedrock Knowledge Bases, and retrieval is filtered by tenant, project and session before anything reaches an agent.
• Each tenant gets its own Kubernetes namespace, Cognito user pool, KMS key, S3 bucket and DynamoDB tables; AgentCore runs each session in its own microVM.

The part worth copying

Self-healing tests are where agent platforms tend to cause trouble, and TCS fenced that step in. The execution agent runs fetch, execute, fix and save as a LangGraph state graph with a fixed order and a bounded retry loop. Only failures classified as code errors go to the fixing agent, which gets a constrained error context. Connectivity and authentication failures do not. Safety checks reject empty files, unexpected new files or extreme changes in file size.

What the post leaves out

There is no cost figure: not per scenario, per tenant or for the model calls. The post does not say which models each agent uses, only that Bedrock lets each agent pick one. It gives no measure of test quality, such as how many generated tests were wrong, how often the fixer's changes were rejected, or whether defects escaped. It also does not compare the platform with a simpler single-agent generator. Engineers still review the output; the post does not say how long review takes or whether it sits inside the 14 minutes.