You're Not the Coder Anymore — You're the Manager: What Alessio Fanelli's Symphony + Linear Setup Means for Designers

OpenAI's Codex-powered agents are now capable enough that at least one builder — Alessio Fanelli, founder of Kernel Labs and co-host of the Latent Space podcast — is running a fully autonomous coding pipeline from his phone, using a combination of OpenAI Symphony and Linear. No laptop required. No sitting and watching a progress bar.

That's the news. The shift it signals is bigger: the job is no longer "write better prompts." It's "manage better systems."

From Prompter to Agent Manager

Fanelli makes a distinction that should immediately reframe how you think about vibe-coding: the difference between being an agent prompter and an agent manager. A prompter fires off a request and waits. A manager sets up a system — defines tasks, tracks state, handles handoffs — and lets it run.

This isn't a personality type. It's a workflow design choice. And it's one designers are actually well-positioned to make, because it's fundamentally about structuring flows and states — not writing code.

The specific setup Fanelli demos uses OpenAI Symphony (an open-source framework from OpenAI for orchestrating multiple agents) paired with Linear (a project management tool most product teams already know). Linear acts as what he calls a "state machine" — a living record of what each agent is doing, what's done, and what's next. Symphony manages the agents themselves through the full development lifecycle. Together, they let him kick off work, walk away, and check back in on his phone.

For designers: think of it less like hiring a freelancer you have to supervise, and more like running a small team with a shared task board. You're still the one deciding what gets built and what "done" looks like. The agents are doing the execution.

What This Setup Actually Looks Like

A few concrete things worth knowing from Fanelli's walkthrough:

Linear as a task board for agents. Linear isn't just a human-readable to-do list here — it's the connective tissue between tasks. Symphony reads from and writes to it, so the agents always know their context. This is meaningful for designers who already use Linear for product work: you might not need a totally new tool, just a new way of structuring issues.

Skills files need a reset, not more rules. Fanelli makes a pointed observation about "CLAUDE.md" files (and the equivalent for other agents) — the instruction files that tell an agent how to behave. His take: most people keep adding rules until the file becomes noise. His recommendation is a full purge and rebuild from scratch. This is counterintuitive but worth sitting with: more instructions don't always equal better behavior.

Token costs are trackable per task. He mentions tracking around 221 million tokens across tasks — which sounds abstract, but the point is that you can measure what autonomous runs are costing you. As these tools move from novelty to workflow, cost-per-task becomes a real variable to watch, not just a vague API bill at the end of the month.

Cloud over local. Fanelli notes that local machines (he mentions Mac Minis specifically) don't scale for running multiple parallel agents. A cloud VPS (a rented remote computer, basically) unlocks the ability to run agents simultaneously without choking your own hardware. Not urgent for small experiments, but worth knowing before you hit a wall.

The Pokémon Cards Demo — and What It Actually Illustrates

The second workflow Fanelli demos is genuinely fun: he uses Codex with browser access to autonomously scout eBay for underpriced Pokémon cards for his San Carlos shop, Merlin Games. The agent browses listings, pulls PSA certificate numbers (authentication IDs for graded cards), cross-references pricing on TCGplayer, and flags deals on cards in the $10K–$20K range.

It sounds like a party trick. It isn't. What it illustrates is that Codex can now browse, extract structured data, and make comparisons — all without a human in the loop. The browser access is the key unlock. An agent that can see the web is fundamentally more capable than one that only works with files you hand it.

For designers building apps: this is the capability that makes "I want my tool to pull in live data" actually achievable without writing a scraper yourself.

What to Watch — and What's Still Unresolved

This is genuinely exciting territory, but a few honest caveats:

Symphony is open-source and early. It exists on GitHub and is real, but like most open-source agent frameworks, expect rough edges and moving parts. It's not a polished product you install in an afternoon.

The "zero babysitting" claim is aspirational for most setups. Fanelli has clearly invested real time in building and tuning this system. The insight isn't "this is plug-and-play" — it's "this is the direction things are moving, and the mental model matters now."

Glimpse gets a mention as a tool that improves "agent senses" (essentially, what an agent can perceive and act on), extending how long autonomous runs can go without needing a human check-in. Worth keeping an eye on as the agent perception layer develops.

The real takeaway: the designer-as-builder role is evolving toward something that looks a lot like systems design. You're defining the workflow, the success criteria, the task structure — and AI handles the implementation. That's not a diminished role. It might actually be the most natural one for designers to step into.