IN Brief:
- TASKING, Infineon, and DLR connected semiconductor information with software implementation, compliance, testing, and verification in one prototype workflow.
- AWS Kiro agents coordinated coding analysis, unit tests, virtual ECU execution, structural coverage, and evidence generation.
- Deterministic verification and engineer approval remain part of each critical decision rather than being delegated to AI.
TASKING, Infineon Technologies, and the German Aerospace Center have demonstrated an AI-assisted engineering workflow that connects semiconductor information, software implementation, compliance, testing, and verification across safety-critical automotive development. The working prototype won the top award at AWS’s Accelerating the V-Cycle with Agentic AI hackathon in Munich.
The project addresses the fragmented toolchains that still characterise much software-defined vehicle development. Requirements, semiconductor documentation, implementation, coding compliance, testing, and verification are frequently handled by different teams and systems, which can leave hardware constraints or verification findings to surface relatively late in the development cycle.
AWS Kiro agents were used to coordinate engineering activity across the workflow. Trusted Infineon device specifications, product updates, and errata were linked with TASKING coding-standard analysis, automated unit-test generation, virtual ECU execution, structural coverage analysis, and evidence reporting.
The TASKING AI Framework allowed the agents to evaluate results and refine code or tests until coding-standard findings, requirements coverage, and structural coverage were either resolved or documented. Engineers retained oversight and approval of critical decisions, while deterministic verification tools remained responsible for the checks on which the resulting engineering evidence depends.
That distinction is central to the prototype. Generative AI can produce code or suggest changes quickly, but safety-critical software cannot be qualified on plausibility. Device behaviour, errata, coding rules, tests, coverage, and evidence need to remain traceable to controlled engineering inputs if the workflow is to support an automotive development process rather than simply accelerate experimentation.
Bringing semiconductor information into the development loop earlier could also reduce the risk of software decisions being made against incomplete hardware assumptions. A device erratum discovered after implementation can force code, tests, and associated safety evidence to be revisited, particularly where the affected peripheral or execution behaviour sits inside a critical function.
The prototype therefore treats AI primarily as an orchestration layer connecting established engineering activities. Infineon supplies the device-level information, TASKING provides compliance and verification tooling, DLR contributes research expertise, and AWS provides the agent environment used to coordinate the workflow. No single AI component becomes the authoritative source for the resulting safety evidence.
The team also designed the proof of concept around cloud-delivered compliance and verification services. That could make specialist engineering tools easier to deploy across distributed development organisations, although production adoption would still require configuration control, tool qualification where applicable, evidence retention, access controls, and integration with the customer’s existing safety lifecycle.
Automotive software programmes are becoming harder to separate cleanly from the semiconductor platforms beneath them. More functionality is being consolidated onto increasingly capable microcontrollers and processors, while software updates, networked architectures, and mixed-criticality workloads increase the amount of evidence required to show that an implementation continues to behave as intended.
AI-assisted tooling can remove manual hand-offs from that process, but the more consequential engineering question is whether it preserves traceability when a requirement, hardware specification, test result, or implementation changes. The TASKING-led prototype provides one architecture for doing that by keeping trusted semiconductor data, deterministic verification, and human approval inside the same loop as the AI agents.
The approach could extend beyond automotive into aerospace, defence, and other safety-critical embedded sectors, where certification similarly depends on auditable engineering evidence rather than the speed at which source code can be generated. Its next test will come outside the hackathon environment, where integration with production requirements systems, configuration management, and long-lived programme baselines will determine whether the workflow can reduce engineering effort without weakening assurance.
