At ATB, we are proud to support the 3D-CIRCULAR project, an innovative European initiative offering…
Research Retrospective: From SmartCLIDE to ASTIR — The Evolution of AI-Supported Software Engineering at ATB
Generative AI has brought intelligent assistance into everyday software development. Developers can now ask an AI system to generate code, explain functions, propose tests or suggest alternative implementations.
Yet one lesson from earlier research remains highly relevant: software engineering involves much more than producing source code.
Between 2020 and 2023, the ATB-coordinated European project SmartCLIDE investigated how intelligent assistance could support software creators across the complete cloud-development lifecycle.
Looking back from today’s generative-AI landscape, several of the problems addressed by SmartCLIDE are becoming more important rather than less.
Beyond code generation
SmartCLIDE developed a cloud-native development environment designed to support the discovery, creation, composition, testing and deployment of data-centred services and applications.
The project deliberately covered several lifecycle stages.
During development, the environment provided abstractions for working with data transformations and processing.
During testing, it supported visualisation of software flows and mechanisms for checking expected behaviour.
During deployment, it abstracted parts of the underlying physical and virtual infrastructure.
At runtime, it provided mechanisms for monitoring service performance and operation.
This broader perspective distinguished the project from tools concerned only with writing code.
Intelligent support before the current GenAI wave
SmartCLIDE was not a large-language-model coding assistant, and it would be inaccurate to present it as one retrospectively.
Its intelligent assistance relied on technologies such as service discovery, recommendation mechanisms, machine learning and higher levels of abstraction.
One of the project’s central ideas was coding by demonstration: rather than requiring every user to describe all technical details manually, the environment could support the creation and composition of services through more guided interactions.
The system was also designed to discover Infrastructure-as-a-Service and Software-as-a-Service resources and help identify suitable infrastructure for an application.
The underlying question, however, is familiar today: how much of the engineering complexity can an intelligent tool take over while leaving developers enough information and control to understand the resulting system?
Making cloud development more accessible
SmartCLIDE also aimed to reduce the barriers to cloud and big-data technologies for SMEs and public organisations.
The project used different levels of abstraction to support users with different technical backgrounds.
An experienced software engineer might work directly with detailed components and configurations. A less specialised user could rely on service discovery, recommendations and higher-level composition mechanisms.
The project was validated through real application scenarios and ultimately delivered a final prototype at Technology Readiness Level 6. CORDIS records the final prototype, pilot assessment material and a dedicated validation procedure among the project’s results.
Why the software lifecycle still matters
Current generative AI systems can produce code much faster than conventional development processes.
But faster code generation can shift rather than eliminate engineering work.
Generated software still needs to fit an architecture. Interfaces must remain consistent. Dependencies need to be controlled. Security and performance must be assessed. Test cases need to cover relevant behaviour. Deployment configurations must be correct, and the resulting system needs to be monitored and maintained.
SmartCLIDE’s lifecycle approach anticipated this problem.
Its architecture treated discovery, development, testing, deployment and operation as connected stages rather than isolated activities.
That lesson is particularly relevant when AI generates increasing portions of an application.
The easier it becomes to produce software artefacts, the more important it becomes to determine whether those artefacts remain coherent as part of a larger system.
From SmartCLIDE to OpenSmartCLIDE
The end of a research project creates another software-engineering problem: what happens to the resulting code?
SmartCLIDE addressed this by transferring its results into the open-source ecosystem through Eclipse OpenSmartCLIDE, based on Eclipse Theia.
At the end of the project, ATB reported that OpenSmartCLIDE included integration with CI/CD environments, intelligent source-code monitoring and dynamic code visualisation. The objective was to continue supporting the software lifecycle beyond the original funded project.
This was not only an exploitation decision. It addressed a fundamental question for research software: long-term usefulness depends on governance, maintainability and integration into existing development ecosystems as much as on the original technical prototype.
What has changed since SmartCLIDE?
The capabilities of AI have advanced dramatically.
Large language models can now generate implementation alternatives, documentation and test cases from natural-language instructions. They can explain unfamiliar code and assist developers across multiple programming languages.
But this creates additional engineering questions.
Where did an AI-generated requirement or implementation decision originate?
Can the generated result be traced back to the engineering information on which it was based?
How can engineers identify plausible-looking but incorrect output?
How should AI-generated artefacts be validated before they enter a production system?
And how can AI tools be integrated into established engineering practices without creating a parallel, opaque development process?
These questions extend the SmartCLIDE challenge from intelligent assistance towards trustworthy AI-assisted engineering.
ASTIR takes the lifecycle perspective into robotics
ATB is now addressing these questions in the Horizon Europe project ASTIR – AI-Augmented Software Engineering for Intelligent Robotics.
ASTIR develops an AI-augmented engineering workbench covering the robotics software lifecycle from requirements engineering to system maintenance. It uses generative AI not only for implementation support, but also for requirements analysis, architectural work, code review, testing and runtime monitoring.
The project completed its first major technology-development phase in June 2026. Initial components now support activities including engineering-document analysis, structured requirements generation, architecture support, software implementation, security review, test generation and runtime monitoring.
The next stage focuses on integrating these components into a common workbench and evaluating them in manufacturing robotics, mobile service robotics and agricultural UAV applications.
A recurring engineering principle
SmartCLIDE and ASTIR use different generations of AI technology and address different application domains.
But they share one important engineering principle.
Useful intelligent assistance cannot stop at producing an artefact.
The artefact must remain connected to requirements, architecture, testing, deployment, operation and maintenance.
This becomes even more important as generative AI takes a larger role in software engineering.
SmartCLIDE explored how intelligent tools could make complex cloud-development processes more accessible. ASTIR now asks how much more capable AI systems can become part of the engineering process without weakening traceability, dependability and professional responsibility.
The technology has changed considerably. The need for disciplined software engineering has not.