How I run autonomous coding agents from my phone with OpenAI Symphony + Linear | Alessio Fanelli (Kernel Labs)

What It Covers

Alessio Fanelli, founder of Kernel Labs, demonstrates how he runs autonomous coding agents using OpenAI Symphony integrated with Linear as a state machine, managing engineering tasks entirely from his phone, and applies the same agentic approach to automate Pokemon card pricing and inventory sourcing for his San Carlos trading card store.

Key Questions Answered

• •Symphony + Linear as agent state machine: OpenAI's open-sourced Symphony framework monitors a Linear board, spins up Codex agents when issues move to "To Do," generates implementation workpads with acceptance criteria, then routes completed work to "Human Review" for PR comments. Moving an issue to "Rework" triggers a line-by-line correction checklist before auto-merging to "Done."
• •Cloud VPS over local runtime: Running coding agents on a cloud VPS (Alessio uses Zillow with 32GB RAM) rather than local machines enables true mobile management. Tasks persist, multiple agents run across four or five projects simultaneously, and you can intervene via Linear, GitHub PR comments, or direct shell access from any device without being physically present.
• •Token cost as a task health metric: Track token usage per task to diagnose agent inefficiency. Most routine tasks run 15,000–60,000 tokens; one Vercel deployment migration consumed 221 million tokens. When actual token spend far exceeds expectations, it signals a gap in tooling, task descriptions, or agent instructions that can be corrected before the next run.
• •Markdown spec files require active maintenance: Agent instruction files accumulate contradictory rules over time because models add lines rather than remove them. Audit and prune these files every few months. Keep specs prescriptive about architecture and commands but avoid over-specifying behavior, since overly detailed skill files cause agents to rigidly follow outdated instructions and create compounding errors.
• •Autonomous web agents for physical-world inventory arbitrage: Codex with browser access can scrape PSA certificate numbers from grading APIs, cross-reference eBay sold listings in batches of five, and flag underpriced cards in real time. The same pattern applies to any inventory business — vintage clothing, fish distribution, trade show card pricing — where heterogeneous data has historically blocked software automation.