Hi guys, my name is Lucas, I'm from Brazil. I am not a programmer. I'm an economist that works in the financial market, but I'm adventuring on the ai coding space to develop personal projects. Did already a (Brazilian tourist focused) vacation destination recommender with AI site mvp, but couldn't advance on the quality of the recommendation algorithm (achieved a score of 6/10 on my evaluation) , then stopped (it was 7 months ago). Now I am starting my second attempt in building a webapp related to agentic e-commerce, also in Brazil. As the majority here are developers, what would you recommend so I could achieve a reasonable mvp? I lack the developer hard skills. I know I couldn't alone build a production grade webapp. But want to test the idea first. I used openhands in jun/25 as I was experimenting with AI coding tools at that time (Cline, roo code, etc. Coding in terminals was not the mainstream) . Came back to openhands only now when I saw the agent canvas, it seems friendlier for a non programmer. I even opened a issue (my first ever issue opened in github) 🙂. @gentle-apple-32922 if openhands is interested in having features/workflows for non programmers in agent canvas, I am opened to help with the testing and workflows concepts. It could reach the mainstream user. Any advice is highly appreciated! I liked a lot the openhands index! One of the best! Since everyone is benchmaxing famous benchmarks, I tend to ignore the majority of them. My current setup is using Droid cli with gpt subscription (via byok). But I wanted to change to a Orchestrator and use also my antigravity subscription (now they have the cli) and maybe codex directly, without Droid in between (I like a lot Droid cli harness for gpt models)

For a non-programmer MVP, I’d make the first version much smaller than the full app idea. For agentic e-commerce, maybe start with one narrow workflow: 1. user describes what they want 2. app asks 2-3 clarifying questions 3. app returns a short list of options with reasons 4. user marks the recommendations good/bad Then keep one simple project doc the agent always reads: - what the app does / does not do - tech stack - test command - important files - known failed attempts - what “good enough MVP” means In my experience, the hard part with coding agents is not only generating code. It’s keeping the product shape stable across sessions, so the agent doesn’t keep rediscovering or changing the same decisions.

Thank you very much! Very insightful to have guardrails so the AI doesn't change what is already working. I am taking a different route. Fixing most demanded 5 types of products and trying to build a database of products with features stored. So the recommendation use this database of features scores. I will implement the project doc as guardrails.

Hi, Lucas oh waveq Welcome 🤝 I don't think we have any roadmap items for non-programmer workflows right now. @hundreds-twilight-32833 can confirm. BUT, if you find a workflow that is working for you, we would love to have you demo at a community call. Our next community call in July 9th. Also, thank you for opening an issue too! I think as you keep using Agent Canvas you'll become more familiar and eventually can work on fixing issues too.

That sounds like a smart route. A product/features database gives the agent something concrete to reason over instead of trying to invent recommendations from scratch. For the project doc, I’d probably include: - the 5 product categories you’re supporting first - what features/scores exist in the database - what the recommender is allowed to do - what it should never change without asking - 2-3 examples of a good recommendation If you end up demoing this on the community call, I’d be interested to see how Agent Canvas handles keeping those rules stable across sessions.

@gentle-apple-32922 thank you for answering. I am still on initial experimentation with agent canvas. I am too used with Droid cli, so I am figuring out how agent canvas could add to my current workflow! Maybe add antigravity cli with Droid cli. I am interested in developing automation/workflows. Didn't succeed on creating custom workflows. A project that potentially had done this really well is open source by Cole Medin Archon, but I found it too difficult for my usage.

That makes sense. If Agent Canvas is still new in your workflow, I’d avoid trying to build a full custom workflow first. Maybe start with a very simple repeatable loop: 1. update the project doc 2. ask the agent to propose the next smallest task 3. have it implement only that task 4. run one check/test 5. write down what changed and what not to change next time For a non-programmer MVP, the workflow matters more than the tool stack at first. If the loop is stable, you can swap Droid / Agent Canvas / Antigravity later without losing the project shape.