Video: "New FREE OpenAI Symphony Update is INSANE!" by Julian Goldie on YouTube.
What Symphony actually is
Symphony is an open-source specification from OpenAI for orchestrating Codex-based coding agents. It is not a model and it is not a new product — it is a structured way of telling agents how to work. Available free on GitHub, it defines how an agent should start a task, which tools it should reach for, how to handle failures, and when to stop and wait for a human review rather than pushing forward blindly.
The practical shape of it is a project-management board — similar to Linear — that acts as a control plane. Each open task gets an agent assigned to it. Agents work continuously. Humans review finished branches. It is designed for software teams that want to run coding agents on real work without babysitting every step.
The agent drift problem Symphony is built to fix
Anyone who has run AI agents on tasks longer than a few steps will recognise the pattern: the agent starts well, works through the obvious steps, and then gradually drifts. It starts solving a slightly different problem, loses track of an earlier constraint, or gets into a loop it cannot escape. The longer the task, the worse this gets.
Symphony addresses this by giving agents a defined operating procedure from the start. The agent does not have to hold its entire remit in one long prompt. It knows which tools to use for specific types of action, how to check its own progress against the task definition, and when a result is complete enough to hand back for review. That structure alone — not a smarter model — is what cuts the failure rate on longer tasks.
Playwright CLI: the update that adds web interaction
The change Julian Goldie focused on is the addition of Playwright CLI as a standard tool in the Symphony setup. Playwright is a browser automation library. With it available, a Symphony agent can now navigate web pages, fill in forms, click buttons, and extract information from sites — all as part of a longer automated job.
In practice this means agents can do things like: check whether a deployed feature works in a real browser, scrape a reference site to gather data for a task, or test a web application after making code changes. Previously you would have needed a separate script or a human to handle the browser steps. Playwright CLI folds that into the agent's normal toolset.
The boot skill: starting right every time
The other piece Julian covered is the boot skill — a short configuration file that tells the agent what it is doing before it starts any task. It sets the tools available, defines what a completed task looks like, and gives the agent a consistent starting state regardless of which project it is working on. Think of it as the operating procedure that a new team member reads before touching anything.
This matters more than it sounds. Without a clear boot state, agents often spend the first part of a task figuring out their own context. The boot skill removes that overhead and reduces the chance of an agent making assumptions that are correct on a simple job but wrong on a complex one.
Where this connects to NordSys
Symphony is the kind of infrastructure change that does not make headlines but does change what is actually possible in production. If you are building automations with Codex or any other AI coding agent, the difference between a structured agent that finishes jobs and an unstructured one that drifts is the difference between something useful and something you have to watch constantly. We help clients design and build AI-powered workflows that actually complete — right through to review and deployment. Our Programming service covers the full stack.
See our Programming service →