Harris Notes

20260507-2007-how-can-agentic-ai--work-with-test-code-written-in

3 min read

how can agentic AI work with test code written in CAPL

Answer

Agentic AI can work with CAPL test code primarily through LLM-powered code generation, analysis, and autonomous test orchestration. The most practical current approach is using GPT-based agents to generate, review, refactor, and explain CAPL scripts, while higher-level agentic frameworks can orchestrate test execution, analyze results, and trigger test modifications in Vector CANalyzer/CANoe environments.

Key Findings

  • GPT-powered CAPL Script Generator: Mohana Krishna Chebrolu (LinkedIn) has built a GPT-based tool that generates, validates, and teaches CAPL scripts covering CAN, CAN FD, Ethernet, SOME/IP, DoIP, and UDS protocols — demonstrating that LLMs can already handle CAPL syntax and automotive protocol logic at a practical level.

  • Agentic AI code-generation pattern: Agents can use LLMs as a ‘reasoning core’ to parse CAPL test specifications, generate new test cases, review existing scripts for errors/edge cases, and refactor code — analogous to how agentic QA tools (Tricentis Tosca, Mabl) autonomously create and maintain test suites in other domains.

  • Orchestration gap: No commercial agentic platform currently offers native CAPL/CANoe integration out-of-the-box. A custom architecture would be needed — e.g., an agent using Vector Scripting API or COM automation to trigger CANoe test runs, parse .asc/.blf log outputs, and feed results back to the LLM for adaptive test generation.

  • Practical architecture: An agentic pipeline for CAPL could combine (1) an LLM agent for CAPL code generation/review, (2) a tool-use layer calling Vector CANoe via its COM API or CAPL DLL interface, (3) a log-parsing module for test result analysis, and (4) a feedback loop to autonomously update test scripts based on failures.

  • Key gap: None of the major agentic test automation vendors (Tricentis, Mabl, Applitools, TestArchitect) explicitly support CAPL or automotive ECU testing environments in their documented product offerings as of late 2025.

Open Questions

  • Does Vector Informatik (CANoe/CANalyzer vendor) have any official AI/LLM integration roadmap or API surface that would enable agentic orchestration of CAPL test execution at scale?

  • Are there automotive-domain-specific agentic AI startups or internal OEM tools (e.g., at BMW, Bosch, Continental) already deploying LLM agents to generate and maintain CAPL regression test suites for HIL/SIL environments?

  • What are the safety/compliance risks of using autonomously generated CAPL code in safety-critical ECU testing (ISO 26262), and what guardrails would be required before agentic AI changes are trusted in production test pipelines?

Entities

twinkle-joshi kaushik-sudhir john-vester mohana-krishna-chebrolu ravikanth-ficusroot testarchitect tricentis-tosca tricentis qtest mabl applitools-visual-ai witbe vector medium linkedin youtube

Concepts

agentic-ai-test-automation capl-script-generation automotive-protocol-testing domain-specific-test-case-generation autonomous-qa-workflow llm-powered-script-assistance human-oversight-in-agentic-qa

Sources

Graph View

Backlinks