I started to build an open-sourced project called Shea Symphony as I was trying to find a way to set up a team of power users of Codex/Clade Code like myself who had already been experienced developers in production as ChatGPT 4 just became available in web chat, and continuously balancing the acceleration impact of AI-assisted coding against the required manageability for delivering software both in terms of speed and quality: it’s much more complicated than having periodic syncing calls because all the tools that we found useful are designed for just individual contributors.
I’ve done loads of research and experiments with existing open-sourced tools but OpenAI Symphony caught much more of my attention: the idea of “moving from managing coding agents to managing work that needs to get done” is pertinent in the first place and the choice of issue tracking systems provided by Linear is intuitive, but what’s surprising to me is that OpenAI released Symphony in the form of a spec document instead of a standalone app or framework which is the typical way of building dependence that can turn into manipulatable market shares.
This is counter-intuitive to me as I’ve got a few vague “principles” based on my explorations on numerous kinds of AI products that I find interesting and potentially valuable, and the first one is that AI products would lose traction if it can foreseeably derived and complemented by OpenAI / Anthropic as they have the biggest motivations and resources to expand their productive scenarios at higher standards (you can easily name a lot of those if you’ve been following), but in the case of Symphony, I’m both surprised in how self-refraining and “over-sharing” that OpenAI was as you get amazed at the quality or just the sheer length of that SPEC.md.
So I reactivated the 200 dollars ChatGPT Pro 20x subscription, and asked Codex to “Make your own” Symphony in Rust, my favourite and production language. Unsurprisingly, Codex didn’t actually bring me what I wished for but what I expected, and I’ve spent around two weeks to make it actually able to run in loops in most of the time and recoverable with manual operations, and a prototype Tauri GUI that can help me built itself “dogfooding”, after reaching #446 in the PR/Issue ticker.
It’s not production ready and there’re still ongoing reiterations on product surfaces, but I’m pretty confident to use it myself as the foundation whenever I’m actually building the team. There are also other implementations of Symphony available on GitHub, and my goal is to make Shea Symphony a better choice for teams that are looking for an opinionated base to extend and build their own.
As I work on Shea Symphony, I’ve been making notes and now I got something to share:
Software ownership can be one of the “moats” against OpenAI / Anthropic in terms of building your products. In the case of 2B products, no organisation will feel safe to build their business fully upon a principally closed ecosystem even if they can achieve better results. The typical compromise is to use commercial models for the actual LLM call, and open-sourced implementations from LangChain/LangGraph/DeepAgents for other nodes.
However, the mindset is very different for 2C products, especially when paywall is involved, users are more willing to grant submission to software if they can achieve what they want without paying extra beyond their preexisting subscription as you can see in the case of OpenClaw, Hermes, and Open Design.
But what’s revealing in OpenAI Symphony is that it seemed to be changing strategy to permeate in the context that marginal benefit of spending more tokens are less game changing and manageable, and more efforts are put to better harness and cost-efficiency in provisioning alternative models.
More than ever before in my opinion, at least within the realm of agentic coding, the narratives of transforming into being more AI-native have been brought back to the reality of human bottlenecks: as models, prompts, context, harness, and loop are eliminating all probabilistic abrasions of human attention, the efficacy of AI-assisted software delivery has started to return to be defined by human in whether developers can properly conceptualise the workflow of creating values by optimising and automating or innovating new mode of production, and providing expertise feedback into the loop to keep it manageable from deviation.
And the way I define Shea Symphony is:
Orchestration and Governance Layer for AI-native Software Delivery: Inspired by OpenAI Symphony, turning tracker-backed engineering tasks into supervised AI-native implementation runs with isolated workspaces, evidence, review, and guarded merge flow.
Extended the Symphony Model from Agent Dispatch into a Full Team Workflow: Configurable workflows, issue forge, team-aware claiming and processing with boundaries, guided human review and approval, reflective backlogging, dreaming, etc.
Manage Work instead of Supervising Coding Agents: Create workflows that let humans focus on intent, structure, and judgment without losing touch with implementation, balancing AI-assisted creative flow with enough engineering harness to keep the code understandable and steerable.
In human words, after setting up Shea Symphony, the desired human workflow looks like this:
Use issue forge skill to discuss your ideas and observation in any agent session to set up issues directly
Use issue forge reflect skill to record sparse ideas into backlogs, collect ideas from previous work, and promote existing backlogs.
Use human review skill for issues waiting for the last UATs before getting approved for merging
Use doctor skill for issues that requires human input to recover.
Other steps are automated programatically for consistency while developers can implement nuanced control by configuring the prompts and models to whatever they like or pick up fully to just work in your favourite coding agent at any stage as you wish.
Hope you like it! You can check it out at https://github.com/Alive24/shea-symphony. At the moment I’m planning a modular/plugin approach to make it customisable within the LangChain ecosystem, and polish it for higher robustness.
PS: I didn’t use any AI to write this post because I’m sharing it like it’s a castle I’ve been building recently for the sheer joy of creating and getting resonance from people who like it, and to some extent easing my anxiety as I look for hiring opportunities.