Why OpenAI Built Symphony and Gave It Away for Free
BBetter Stack
Computing/SoftwareManagementInternet Technology
Transcript
00:00:00This is OpenAI's Symphony, an open-source tool for orchestrating long-running agents using an
00:00:05existing issue tracker like Linear to help your agent autonomously complete tasks without any
00:00:10human supervision. But why doesn't an agent have to build it from scratch before you can use it?
00:00:14Does it only support Codex CLI? And is this the beginning of more
00:00:18open-source tools from OpenAI? Hit subscribe and let's find out.
00:00:25Symphony only exists because OpenAI hit a human-attention bottleneck, which means engineers
00:00:30could only supervise 3-5 Codex sessions concurrently before context switching began
00:00:35to negatively affect productivity. This, of course, wasn't going to scale. So guess how OpenAI fixed
00:00:41the fast-agent-slow-human-managers problem? They got rid of the human-managers. Kind of.
00:00:47Because with Symphony, humans put tasks on a board, a new agent will be spun up to complete that task,
00:00:52and the agent will only involve the human if there's something to review.
00:00:55But how does Symphony compare to similar tools like Multica and Conductor? Well,
00:00:58watch the demo and that'll be made crystal clear. Before we start, however, I just want to say
00:01:03Symphony has the weirdest installation process I've ever seen in my life. Or it could be the
00:01:07most genius way to install something. We'll get to that later.
00:01:10First, however, let's go through a basic example. So I have Symphony running,
00:01:14polling for tasks in Linear to work on. And in Linear, I'm going to create a new issue,
00:01:18which will be to build a Hello World app using TypeScript and BUN.
00:01:22Right now, Symphony isn't configured to work on backlog tasks. So I'll change the status to To-Do
00:01:27and hit Create Issue. Remember the ID of the task, which is SYN7.
00:01:31So after a while, Symphony picks up on the ID of that task. And a few seconds later,
00:01:36we have a GraphQL validation error, which isn't a big deal and is very easy for Codex to fix.
00:01:41But after that, we can see that Codex has completed the work, has changed the issue
00:01:45from To-Do to Done, and has left a comment here from Symphony. Which means if we go to
00:01:49our Symphony workspaces directory, we'll talk more about that later, we can see we have a new
00:01:53workspace with the same ID as our issue. And if we go into that workspace, we have the list
00:01:58of files created for the Hello World TypeScript BUN app. And if we go into the source directory,
00:02:04we can see that the code for the app is over here. And that's essentially Symphony in a nutshell.
00:02:08Now let's go through how to set it up. So there are two ways to install Symphony
00:02:12according to this repo. Option two, which is what we're used to, we set up Elixir,
00:02:16clone the repo, then build the code and run it using the existing workflow file.
00:02:20Option one, however, is possibly the weirdest or most forward thinking way to install something.
00:02:25You just basically give your coding agent this prompt, and it will read the spec file,
00:02:30which is more than 2000 lines long. But it basically gives your agent detailed
00:02:34instructions on how to build Symphony, which is wild because if everyone went down this route,
00:02:39no two versions of Symphony would look the same. Some people will have different features in
00:02:43different languages, which would be chaos for open AI to maintain and support.
00:02:47But it's also kind of genius because if you built your own version of Symphony,
00:02:51you would feel responsible for it. You'd fix the bugs, you'd add the features,
00:02:55and you'd essentially maintain it. And if you didn't want Symphony to use linear or
00:02:59codecs, then that's up to you. Someone built a Go version of Symphony that runs on Charm CLI,
00:03:04and someone else built one that was powered by the Claude SDK. I wasn't that creative,
00:03:09so I put the default prompt into codecs using GPT 5.5 Low Effort, and it gave me a Python version
00:03:15of Symphony, which makes sense because NLM's are very good at Python. But once that's done,
00:03:19you'll need a linear personal API key, which you can get from security and access,
00:03:23and you can create one by clicking over here. Then you'll have to add that key to your work profile,
00:03:28which is a file that tells Symphony how to do its job, containing some YAML front matter,
00:03:32which of course has the API key, active states for the agent to know when it can work on a task,
00:03:37as well as the workspace root and the codecs command, which is the shell command Symphony
00:03:42uses to launch a coding agent. And below that is the markdown prompt sent to the coding agent
00:03:46for each issue. You can access OpenAI's workflow in D file from their repo if you're interested.
00:03:51But as of right now, this workflow isn't fit for a real project. Say I wanted to make changes to
00:03:56my film emulation app that I've been working on. I would need to add a create-after hook, which runs
00:04:01after Symphony creates a workspace for an issue, and this hook will first clone the repo to the
00:04:06workspace directory, and then check out to a new branch in that repo. I've also added a run-after
00:04:10hook, which runs after codecs finishes working on the issue. So this one stages the files, then
00:04:15creates a new commit before pushing the branch and creating a new pull request with these values.
00:04:20So now if I run Symphony using UV, which was also acquired by OpenAI, but that's a topic for another
00:04:25video, it's polling for new issues. So now if I create a new issue to update the readme, just in
00:04:30the section that shows batch processing, to use a wildcard instead of showing multiple files, again,
00:04:35I'll change it from backlog to to-do, then hit create issue. Now Symphony has picked up on that
00:04:40issue. And if I look at the Symphony workspaces directory, we can see there's a new folder matching
00:04:45our issue ID, which contains the cloned project. And now that Symphony has finished working on the
00:04:49issue, it's changed the status to done and has created a new PR with a link to the linear ticket
00:04:54and the exact change that I requested. Now, of course, I could change my Symphony code to add a
00:04:59better PR description and to put a link to the PR in the comments, but I'm sure that's very easy for
00:05:04Codex to do. So that's a quick overview of Symphony. If you're already comfortable with or work at a
00:05:08company that uses linear, then you'll feel right at home with the interface. And if you're a Codex user,
00:05:13then you can use your skills, MCP tools and plugins that you already have installed. I personally don't
00:05:18use Codex, but I can definitely see the vision OpenAI we're going for with Symphony. Imagine
00:05:22you had a team of developers all working on the same project with AI. Instead of having their own
00:05:28personal harness with their own workflow, there can be a central agent with central skills, central
00:05:33tools, central workflows and plugins. And each developer can communicate with that by giving
00:05:37it a task and they can see what other developers are working on, what prompts they've used and can
00:05:41see at a glance what other features might affect the feature that they're working on. Even though
00:05:46it's very difficult to come up with a feature that won't touch another part of the code that someone
00:05:49else is working on. But as cool as Symphony is, in my opinion, Multica does a better job because it's
00:05:54easier to set up, create and add different agents, and you can even schedule tasks. Nevertheless,
00:05:59I completely see a place for Symphony. It's kind of like the bare bones version of Multica where you
00:06:04can add the exact features you want instead of having a product that is built for you. So think
00:06:08of Symphony like the Pi harness and other tools like Multica or Conductor as cloth code. And if
00:06:13you don't know anything about Multica, then check out this video, which goes through everything you
00:06:18need to know about it.