WITNESS — From Intent to Autonomous Execution

The Problem

AI agents can generate code and suggest solutions, but getting useful work done requires more than generating an answer. Tasks must move through defined steps, automation must respect its boundaries, and developers need a way to inspect what actually happened.

Without a clear workflow, it becomes difficult to distinguish an AI-generated suggestion from an action that was genuinely executed.

Our Solution

WITNESS explores a more structured approach to AI-powered software development automation using the GitLab Duo Agent Platform.

The project implements three automation flows—Assisted, Supervised, and Hands-off—alongside a dedicated agent. Together, they demonstrate different levels of automation and human involvement within a defined workflow.

Assisted keeps the developer closely involved while the agent supports individual tasks.

Supervised supports a broader automated process while retaining human oversight.

Hands-off demonstrates a more autonomous execution path through explicitly configured workflow components.

Rather than treating autonomy as a simple on/off switch, the project makes different automation approaches inspectable through their underlying flow configurations.

How It Works

The implementation contains three distinct flows with verified component counts:

• Assisted: 6 components
• Supervised: 7 components
• Hands-off: 8 components

Automated tests check these counts against the actual flow configuration files, helping prevent discrepancies between the implementation and its documentation.

The project also includes automated testing and GitLab CI/CD integration. Its existing pipeline history provides evidence of a successful run with 12 of 12 jobs passing across nine stages.

Why It Matters

The goal is to make AI-powered automation easier to understand, evaluate, and trust.

By separating assisted, supervised, and hands-off execution into explicit workflows, WITNESS provides a concrete demonstration of how different levels of developer involvement can be represented within the GitLab Duo Agent Platform.

The central principle is simple: automation should be judged by the work it actually performs, not merely by what an AI claims it can do.

Implementation and Verification

The project is publicly available on GitLab, with its workflow configurations, documentation, and automated tests.

The accompanying demonstration video walks through the implementation and illustrates the three automation flows.

The result is a practical exploration of structured AI workflow automation, with an emphasis on implementation evidence, consistency, and verifiability.

Built With

• agent
• agents
• ai
• automated
• automation
• ci/cd
• duo
• gitlab
• platform
• python
• yaml