Summary
Alessio Finelli, co-founder of Kernel Labs and co-host of the Latent Space podcast, joins Claire Vo on How I AI to demonstrate how he orchestrates autonomous coding agents from his phone using OpenAI’s Symphony framework combined with Linear as a task management layer. Rather than running agents locally, Finelli moved his entire setup to VPS instances in the cloud, enabling him to monitor and interact with long-running agents via text message, the Linear interface, or direct shell access — all from mobile.
The interview walks through real use cases from Finelli’s businesses: automating inventory tracking for a fish delivery company using computer vision, building a Pokémon card pricing tool with OpenAI Codex that looks up values in real time at trade shows, and managing engineering tasks for his game store Merlin Games in San Carlos. He explains how Symphony instances are structured with named workspaces, task logs, token usage tracking, and state files, and how the shift from being an “agent prompter” to an “agent manager” changed both his architecture and his results.
A key practical insight concerns agent memory files like CLAUDE.md: Finelli argues these accumulate stale and contradictory instructions over time because models tend to add rather than remove directives. He recommends periodic audits to prune them, warning that overly detailed skill files can make models rigidly follow outdated rules. The episode is sponsored by Firecrawl and Jira Product Discovery.
📺 Source: How I AI · Published July 06, 2026
🏷️ Format: Workflow Case Study
