Submind YouTube summaries
Thumbnail for Meet the Flow Creator Agent in GitLab 19.3: Prompt-to-Workflow Demo

Meet the Flow Creator Agent in GitLab 19.3: Prompt-to-Workflow Demo

Watch on YouTube

Video summary

The Flow Creator Agent, introduced in GitLab 19.3 as a foundational agent, simplifies the creation of custom automation flows by allowing users to describe desired workflows in plain English rather than learning complex configuration schemas. Previously, setting up automated processes required understanding the flow registry schema, but this new agent bridges that gap by generating complete, runnable flow definitions directly from natural language prompts. The demonstration begins with a practical scenario involving Tanuki Bank's fraud detection microservice, where an open merge request contains a conflict that needs resolution. Instead of manually navigating technical settings, the presenter uses the agent to design a triage flow that automatically detects conflicts, adds necessary labels, and notifies the author with suggested next steps, effectively handling the initial identification phase without requiring any code changes. The agent's capabilities extend beyond simple detection by actively researching available tools within the GitLab environment to ensure the proposed workflow is feasible. During the initial design phase, the Flow Creator Agent identifies that a dedicated label tool is missing from the current set and creatively utilizes an existing quick action in merge request comments to achieve the same result. Once the first version of the flow is created and enabled with a specific trigger for merge conflict events, it runs alongside other automations like Duo Code Review on the same merge request. This highlights how multiple agents can collaborate on a single project, with the Flow Creator Agent successfully checking conflicts, applying labels, and tagging authors to streamline the review process before any manual intervention is required. Recognizing that identification alone does not fully resolve the issue, the presenter iterates on the workflow by asking the agent to evolve from a triage tool into an active conflict resolver. The agent responds by suggesting significant configuration changes, such as incorporating Git and file tools to perform actual code modifications, splitting the process into distinct agentic steps for fixing conflicts and posting summaries, and adding testing capabilities. This updated version instructs the agent to check out the source branch, run a git merge against the target, resolve the conflict automatically, execute pytest tests, and push the resulting commits back to the original branch. The result is a fully automated solution that not only identifies the problem but also fixes it, runs necessary tests, and reports success, leaving the merge request ready to be merged without any remaining conflicts. Ultimately, the Flow Creator Agent empowers teams to build robust automation by starting small with simple natural language instructions and progressively enhancing capabilities based on real-world needs. The iterative approach allows users to see what is missing in their current setup and continue refining the workflow until it perfectly aligns with how their team operates. By combining ease of use with powerful agentic reasoning, this tool transforms complex multi-step tasks into seamless automations that adapt over time. As part of the GitLab 19.3 release, the Flow Creator Agent represents a significant step forward in making advanced automation accessible to everyone, enabling developers to focus on solving problems rather than configuring intricate schemas.
Read the full video transcript
Hi, hi. It's Fatima. Today, I'm excited to show you Flow Creator Agent, one of our newest foundational agents in GitLab 19.3. On GitLab, custom flows can already turn manual multi-step work into automation. [music] To write a flow though, you had to understand the flow registry schema first. [music] Flow Creator Agent lets me skip that part. I can describe the automation I want in plain English and it gives me a complete runnable flow definition that I can register into the AI catalog. So, let's try this on one of my favorite problems, a merge conflict. This is Tanuki Bank's fraud detection microservice. It evaluates transactions in real time and returns a risk score, risk level, and recommendation. Now, there's an open merge request here to update the calculate route and you guessed it, it has a merge conflict. I know exactly how I want us to handle this, but I really don't want to learn a whole configuration schema before I can set up a flow. So, today I'm going to deal with this using the Flow Creator Agent. I'm going to ask for a simple merge conflict triage flow. When a conflict is detected, I want it to check that the conflict is still there, add a needs resolution label, and let the author know in a useful comment with next steps. For now, no code changes, just a flow that does triage. Now, Flow Creator Agent checks the current flow registry docs before it designs the flow and starts to do some research. So, let's see what it finds. It found the merge request trigger with the merge conflict action and it noticed that there isn't a dedicated label tool in this tool set. So, it uses the {slash} label quick action in the merge request comment instead. It also gives me a quick checklist to review before I register it. Now, I'm happy with this first version, so I'll go ahead and create a new flow. I'll head to AI and then flows. I'll call this Tanuki Bank merge conflict and add a description. I'll set it to public and copy the configuration from the sidebar where Flow Creator Agent made it for us into the flow editor. Once the definition is saved, I can enable the flow and will need to pick an event trigger. There are a lot of events that I could use here from mention to assignment, but for this one I want a merge request event and specifically the merge conflict action. That means I don't have to be watching the merge request and manually kick this agent off when the conflict appears. Now, I'll enable the flow and there's the service account. This is the composite identity for the flow. So, when this agent operates, this is the identity it uses when it's working in the project. Now, let's go back to the merge request and see what happened. We can see two sessions running here. One is the catalog agent, that's our flow, and the other is duo code review, which was triggered because the merge request became ready. This is a great example of how you can have multiple automations working on the same merge request because teams will usually have more than one thing they want to happen when their code is ready for review. And here's the comment from our flow. It's checking the conflict, adding the label, suggesting next steps, and tagging the author. Now, this is useful, but it's also pretty simple. It tells me what to do, but I still have to do the manual work. So, naturally I'm thinking, what if the flow could also resolve the conflict for me? Let's go back to the flow and open up Flow Creator Agent again. This time I want to turn merge conflict triage into merge conflict fix there. We can keep the merge conflict trigger, summarize what's wrong, and then use git merge to resolve it, run the test, and push the fix to the source branch, ending with a summary for the author to review. This is the part that I really like about Flow Creator Agent. I'm not trying to design the perfect automation in one shot, but I can start with the process that I understand and then continue to iterate on the capabilities as I know what my team needs. So, as the Flow Creator agent is researching, it gives us some important configuration changes. Because this version of the flow will work with files and push a commit, it now needs Git and file tools to do the actual work. It also suggested to split this into two agentic steps. One component that fixes the conflict and a second one that posts a summary back to the merge request. In this version, it'll check out the source branch, run Git merge against the target, resolve the conflict, run the Pytest test, and then commit and push those results to the original source branch. I'll copy over this updated configuration, save it into the flow, and head back to the merge request. And this time, the flow has actually resolved the conflict and pushed commits onto the source branch. I can review the changes, check the results, and make sure that the merge was done successfully. And when I flip back, the merge request is now ready to merge. No more conflicts. We started with a merge conflict and a simple flow that identified the conflict. Then we iterated on it so that it could make the change, run the [music] test, push the fix, and report back. With Flow Creator agent, you can start small with natural language, see what's missing, and keep building until the automation fits [music] the way that your team works. Flow Creator agent will be available with the 19.3 release. Oh, this just gave me an idea for a flow that I need for my team. So, I'm going to go work on that, [music] and I'll see you next time. Bye. Thanks for watching.