I Tried an AI Coding Tool Built Like a Delivery Team (Routa)

BBetter Stack
컴퓨터/소프트웨어경영/리더십AI/미래기술

Transcript

00:00:00This is Ruda, an open-source AI coding tool
00:00:03that turns your agents into something closer
00:00:05to a delivery system.
00:00:07Not another paste your repo context and pray it works,
00:00:11a delivery system with backlogs, dev, review,
00:00:14evidence and gates.
00:00:16Because AI coding tools made us faster,
00:00:19but they also gave us a new job, managing the AI.
00:00:22Ruda is free, local first and built
00:00:25around a Kanban board for AI agent.
00:00:27Let's see if this actually holds up.
00:00:30(logo dings)
00:00:34Most AI dev tools hit the same three walls.
00:00:36First is chat hell.
00:00:38Everything important lives inside a conversation.
00:00:42The plan, that failed attempt, the fix,
00:00:44that weird work around, it's all trapped us
00:00:46just scrolling back.
00:00:47We do it all the time.
00:00:48Then there's no traceability.
00:00:50The AI changes code, but you don't always know
00:00:53what it tried, why it chose that approach,
00:00:56or what evidence it actually used to back it.
00:00:58Then finally, there's no real quality gates.
00:01:01We still have to ask things like, did it run the tests?
00:01:04Did it check the diff?
00:01:06Did it follow the actual acceptance criteria?
00:01:08And that's the thing here, because writing code
00:01:10is not the same thing as delivering software.
00:01:13Ruda's idea is kind of simple here.
00:01:16Stop treating AI coding like a chat session.
00:01:19Treat it like a delivery pipeline with tasks,
00:01:22agents, review stages, evidence and gates.
00:01:25Basically, CI or CD thinking for AI assisted software work.
00:01:30Now, watch how that changes the workflow.
00:01:32If you enjoy coding tools to speed up your workflow,
00:01:34be sure to subscribe.
00:01:35We have videos coming out all the time.
00:01:37This is the Ruda desktop app.
00:01:39It can be self-hosted with Docker just by pulling the repo
00:01:42and running Docker Compose up for all you self-hosters.
00:01:45Though I opted for the desktop after some issues
00:01:48trying to sync my Git repo, so it's a plug and play there.
00:01:52I'm gonna create a workspace, attach a real repo I have
00:01:56and choose Kanban.
00:01:58Then it opens this up and I'm gonna give it
00:02:00one or two small tasks.
00:02:02Nothing dramatic guys, just I don't wanna build
00:02:05the whole app, just the kind of task you would actually
00:02:07hand to an AI tool during normal development.
00:02:11Normally, this is where I would open a blank chat
00:02:14and try and write a really good prompt.
00:02:16But here, after I add the task, it automatically goes through
00:02:19and adds it to the Kanban board.
00:02:22The task is now not floating around in my chat.
00:02:25It now actually has a place to go.
00:02:27And it's gonna start right here in the backlog.
00:02:30As the AI progresses, it moves it into development
00:02:34and the right agent is gonna pick up on them.
00:02:36Now you can see the handoff and all this
00:02:38is being done automatically.
00:02:39So it's going through the different stages here.
00:02:41Now that sounds small, but it matters.
00:02:44I can still check the flow, the output
00:02:46and what the AI is doing.
00:02:48Within a task, you could also even chat with it
00:02:51while it's running the process.
00:02:54So instead of one giant conversation pretending
00:02:56to be a workflow, this workflow is now visible.
00:02:59I get evidence, I got traces.
00:03:01I can see what actually changed, what was checked
00:03:03and where the task is within the process of development.
00:03:07This uses open code and any other AI agent you really want.
00:03:11You can connect your AI API and just choose the one you want.
00:03:14For all this, I synced up my anthropic key to use Claude.
00:03:18Now let's break down what Ruta actually is.
00:03:21It's not trying to be another AI chat box.
00:03:23That's the key thing to understand here.
00:03:25The main idea really is the Kanban board.
00:03:28It's the coordination layer.
00:03:30Think of it like the project board
00:03:32your AI agents have to work through.
00:03:35A task starts in one lane, it moves to another
00:03:37and it goes through the lanes.
00:03:39Different agents can handle different stages
00:03:42if you link them up with your AI keys.
00:03:44So instead of one agent trying to do everything,
00:03:47plan the work, write the code, review, test, explain,
00:03:50Ruta gives the work structure.
00:03:52You create a workspace, you connect a repo,
00:03:55you define a task.
00:03:56The agents work inside of that structure.
00:03:59It also uses agent protocols like MCP and ACP.
00:04:03So you can add those in or use those where you need.
00:04:06It's more like infrastructure
00:04:08for coordinating software agents.
00:04:10And Ruta is not just asking an LLM,
00:04:12"Hey, Claude, does this look good?"
00:04:14It's trying to add checks, fitness functions,
00:04:17evidence, review gates.
00:04:19That changes the question, right?
00:04:22How are we using this?
00:04:23How can this speed up our workflow?
00:04:25Now, a lot of devs go through the same AI coding curve.
00:04:29You ask for a function, it spits one out.
00:04:32You ask for tests, it spits those out too.
00:04:34You paste an error, what does it do?
00:04:36It gives you a fix.
00:04:38Well, hopefully it gives you a fix.
00:04:41But then you start using it on a real code base.
00:04:43And that's when it all turns into actual maintenance.
00:04:46The context gets messy,
00:04:48the agent forgets what it already tried.
00:04:51It changes files you didn't even ask it to touch.
00:04:53You have to keep checking its work.
00:04:55And what's happening here is little by little,
00:04:58something annoying is happening.
00:05:00You didn't get rid of the work.
00:05:02You are now just the manager for the AI.
00:05:05It's a different kind of work.
00:05:06We're tracking the tasks, we're reviewing the diff,
00:05:09we're checking the tests, all that stuff.
00:05:12So it's just a new job.
00:05:13Ruta gives you a visible board for all of this.
00:05:17So where does this even fit compared to tools
00:05:19we're already using, right?
00:05:20Tools like Cursor and Clod, they're chat first,
00:05:23which is not a bad thing, right?
00:05:25They're great when you want a really strong code assistant
00:05:28that's really integrated and close to our code.
00:05:31But the center of gravity is still the conversation,
00:05:34the prompts we're giving it.
00:05:36Ruta's center is a bit different.
00:05:38It's the task moving through a delivery system,
00:05:41backlog to do testing review.
00:05:44Now compare that to agent frameworks like CrewAI
00:05:46or LaneGraph, those are more flexible.
00:05:49But that flexibility means you often have
00:05:51to build the workflow yourself.
00:05:52Who plans, who implements, where does the evidence go?
00:05:55All that stuff.
00:05:56Ruta is free, local first, and it's pluggable.
00:05:59Local repos, local workflow, there's no mandatory account
00:06:02just to try the basic idea.
00:06:05This is not where I say drop everything,
00:06:08replace it with Ruta.
00:06:09No, come on, I'm not gonna do that.
00:06:10That's not true at all.
00:06:12And this is only good for, well, certain things.
00:06:15I had a good time with it.
00:06:17There are some things that I liked right away.
00:06:19A visible board, clear task states,
00:06:22and traceable handoff, that's great.
00:06:24That's more professional than a 300 message chat thread.
00:06:27The local first approach is great.
00:06:30A lot of us are tired of AI tools becoming
00:06:32a subscription model with no boundaries around our code.
00:06:35Being able to keep this close to our local workflow
00:06:38is a real advantage.
00:06:40But I mean, yeah, technically you could do
00:06:42all this on your own, but this does not help
00:06:44keep things more organized, so in my mind,
00:06:47using Ruta actually speeds up the flow.
00:06:49The Kanban and protocol model has a learning curve.
00:06:52And if you just wanna open a chat box
00:06:53and ask a question, paste an answer,
00:06:56this is too much, right?
00:06:57That's not what this is at all.
00:06:59The desktop app is the best way to use it,
00:07:01but it's not gonna feel as great
00:07:02as the biggest commercial AI.
00:07:05I mean, Cursor feels good.
00:07:06It's that interface, right?
00:07:07Clawed code.
00:07:08And there are fewer ready-made agents
00:07:10that you might get from closed tools.
00:07:12But that's actually kind of why I like the direction.
00:07:15It's not pretending the hard parts
00:07:17of software delivery disappeared.
00:07:19It's just trying to organize them.
00:07:20AI is not going away, but the chat-first workflow
00:07:22is starting to show its limits.
00:07:24And the next step is not just smarter models,
00:07:27it's better coordination, it's better traces,
00:07:29it's better gates.
00:07:30If you enjoy coding tools like this,
00:07:32be sure to subscribe to the Better Stack channel.
00:07:34We'll see you in another video.

Key Takeaway

Routa shifts AI-assisted software development from a disorganized chat-based process to a structured delivery pipeline using a Kanban board to manage tasks, agent handoffs, and quality gates.

Highlights

  • Routa manages AI agent workflows using a Kanban-style board rather than a traditional chat-based interface.

  • The tool addresses traceability issues by tracking agent progress, output, and modifications through defined delivery stages.

  • Routa functions as a local-first application that integrates with personal Git repositories via Docker Compose or a desktop app.

  • Developers can connect custom AI API keys, such as Anthropic's Claude, to power the agents within the infrastructure.

  • The system enforces quality control by allowing users to implement check gates and fitness functions within the development pipeline.

  • Routa distinguishes itself from chat-first tools like Cursor by shifting the center of gravity from prompt interaction to task-based delivery.

Timeline

Limitations of chat-based AI development

  • AI coding tools often trap plans, failed attempts, and fixes within unstructured conversation threads.
  • Chat-first interfaces lack inherent traceability regarding why an AI chose a specific approach or what evidence it used.
  • Writing code through standard chat interfaces misses necessary quality gates like test verification and acceptance criteria checks.

Standard AI development tools rely on chat sessions that hide critical development context and decision-making logic. This approach creates a new burden on developers who must act as managers for the AI rather than just coding. Effective software delivery requires more than just generating code; it necessitates a structured pipeline with tasks, review stages, and evidence tracking.

Implementing a delivery-focused workflow

  • The Routa desktop app provides a self-hostable environment that can be deployed via Docker Compose.
  • Workspaces connect directly to local Git repositories and visualize task progress on a Kanban board.
  • Tasks move automatically from backlog to development and review, providing visibility into the AI's actions.
  • Users maintain the ability to chat with the agent within a specific task context while it executes the process.

Routa treats AI coding as a delivery system similar to CI/CD pipelines. Users define tasks that move through specific lanes, ensuring that the work is visible and organized. This framework allows for automatic handoffs between stages and provides clear traces of what changed in the code, what was checked, and the current status of each development task.

Architecture and comparative positioning

  • Routa acts as a coordination layer for AI agents, allowing different models to handle different stages of the delivery process.
  • The infrastructure supports agent protocols like MCP and ACP to manage software agents effectively.
  • Unlike chat-first tools like Cursor, Routa prioritizes task movement and structured evidence gathering.
  • The system imposes a steeper learning curve compared to simple chat boxes due to its Kanban and protocol-based model.

The tool functions as infrastructure rather than just another interface for an LLM. It focuses on organizing the hard parts of software delivery—such as planning, writing, testing, and reviewing—rather than attempting to replace the AI model itself. While it lacks the polished interface of commercial tools, it offers superior organization and transparency for developers who need to manage complex, multi-agent workflows locally.

Community Posts

No posts yet. Be the first to write about this video!

Write about this video