How I run autonomous coding agents from my phone with OpenAI Symphony + Linear | Alessio Fanelli (Kernel Labs)
Alessio Fanelli demonstrates how to run autonomous coding agents from a phone using OpenAI's Symphony framework integrated with Linear, showing practical applications from Pokemon card arbitrage to small business automation. He emphasizes moving from agent prompting to agent management through cloud-based orchestration and shares how AI enables efficient scaling of traditionally manual, inefficient business processes.
Summary
Alessio Fanelli, founder of Kernel Labs, discusses his approach to autonomous coding agent management using OpenAI's Symphony framework combined with Linear project management. Rather than running agents locally, he hosts everything on a VPS in the cloud, allowing him to manage tasks via text, Linear integration, or direct shell prompting—enabling work from anywhere, including his phone.
The core system uses Symphony as an orchestration loop that converts Linear issues into coding tasks. When a task moves from 'To Do' to 'In Progress' status, Symphony creates a Codex workpad where the agent plans implementation with acceptance criteria and validations. The workflow.md file serves as the spec explaining what the agent should do. After completion, tasks move to 'Human Review' where code can be commented on GitHub before either moving to 'Done' or 'Rework' with a checklist addressing all feedback.
Fanelli emphasizes that Symphony is not about new capabilities but about shaping context effectively. He tracks token usage per task, discovering that understanding expected token costs helps identify inefficiencies in tooling or specifications. Most tasks cost 15-60 million tokens, though complex infrastructure changes can reach 221 million tokens. This data informs future optimizations.
Beyond coding orchestration, Fanelli showcases the Pokemon card arbitrage business at his San Carlos store, Merlion Games. He uses Codex with two main skills: one retrieves PSA grading certificates from the internet (requiring browser access to extract data from images), and another finds underpriced cards on eBay by comparing current prices against his inventory database. The system batches searches in groups of five to avoid being flagged, accounts for different grading equivalencies (PSA 10 vs BGS 9, etc.), and identifies cards worth purchasing in real-time at trade shows.
Fanelli stresses that having clear, minimal skill.md files works better than verbose ones—models tend to add rather than remove instructions, causing confusion. He recommends regularly purging markdown files as models improve. He also notes that AI's value in long-running tasks isn't primarily response time but the ability to save clock time for real people by automating heterogeneous, data-rich decisions.
He sees AI's most positive outcome as enabling small business creation and efficiency at scale, citing examples like his father's fish delivery business (inventory tracking), vintage clothing reselling, and his own trading card operations. The technology allows businesses with limited human bandwidth to capture previously inaccessible market opportunities. He also discusses personal uses like offloading financial decision-making stress and email triage to AI as a safety net.
About this episode
<p><strong>Alessio Fanelli</strong>, founder of Kernel Labs and co-host of Latent Space podcast, walks us through two very different AI workflows: (1) a fully autonomous coding setup using OpenAI Symphony + Linear, where Linear acts as a state machine and Symphony manages agents through the whole dev lifecycle with zero babysitting; (2) Codex with browser access searching eBay for underpriced Pokémon cards—autonomously browsing, extracting PSA certificate numbers, and flagging deals on $10K–$20K cards for his San Carlos card shop, Merlin Games.</p><p><br /></p><p><strong>What you’ll learn:</strong></p><ol><li>Why “agent manager” is a better mental model than “agent prompter”</li><li>Why local Mac Minis don’t scale, and what a cloud VPS unlocks</li><li>How to wire Symphony and Linear together as an agent state machine</li><li>How to track token costs per task (and what 221 million tokens buys you)</li><li>What Glimpse does, and why better agent senses extend autonomous runs</li><li>Why your CLAUDE.md probably needs a full purge, not more instructions</li><li>How Codex scouts underpriced $10K Pokémon cards on eBay at scale</li><li>The new category of small business that AI just made possible</li></ol><p>—</p><p><strong>Brought to you by:</strong></p><p><a href="https://firecrawl.dev/?utm_source=newsletter&utm_medium=partner&utm_campaign=how_i_ai" rel="ugc noopener noreferrer" target="_blank"><strong>Firecrawl</strong></a>—Power AI agents with clean web data</p><p><a href="https://atlassian.com/howiai" rel="ugc noopener noreferrer" target="_blank"><strong>Jira Product Discovery</strong></a>—Prioritize with insights, build with confidence</p><p>—</p><p><strong>In this episode, we cover:</strong></p><p>(00:00) Intro</p><p>(02:24) Prompter vs. agent manager</p><p>(04:31) Live demo: Symphony + Linear</p><p>(09:31) Setting up Symphony</p><p>(14:15) Purging your skills files</p><p>(18:06) The benefits of this system</p><p>(19:10) Demo: Using Codex to hunt for Pokémon cards</p><p>(24:17) The benefit of AI for small businesses</p><p>(28:23) Lightning round</p><p>—</p><p><strong>Tools referenced:</strong></p><p>• OpenAI Codex: <a href="https://openai.com/codex" rel="ugc noopener noreferrer" target="_blank">https://openai.com/codex</a></p><p>• OpenAI Symphony (open-source framework): <a href="https://github.com/openai/symphony" rel="ugc noopener noreferrer" target="_blank">https://github.com/openai/symphony</a></p><p>• Linear (project management/agent state machine): <a href="https://linear.app/" rel="ugc noopener noreferrer" target="_blank">https://linear.app</a></p><p>• PSA (Professional Sports Authenticator) grading: <a href="https://www.psacard.com/" rel="ugc noopener noreferrer" target="_blank">https://www.psacard.com</a></p><p>• TCGplayer (card pricing): <a href="https://www.tcgplayer.com/" rel="ugc noopener noreferrer" target="_blank">https://www.tcgplayer.com</a></p><p>• eBay (used for card price scouting): <a href="https://www.ebay.com/" rel="ugc noopener noreferrer" target="_blank">https://www.ebay.com</a></p><p>—</p><p><strong>Other references:</strong></p><p>• Meta Ray-Ban glasses: <a href="https://www.ray-ban.com/usa/ray-ban-meta-smart-glasses" rel="ugc noopener noreferrer" target="_blank">https://www.ray-ban.com/usa/ray-ban-meta-smart-glasses</a></p><p>• <em>The Monk and the Riddle</em> by Randy Komisar: <a href="https://www.amazon.com/Monk-Riddle-Creating-Making-Living/dp/1578516447/ref=sr_1_1" rel="ugc noopener noreferrer" target="_blank">https://www.amazon.com/Monk-Riddle-Creating-Making-Living/dp/1578516447/ref=sr_1_1</a></p><p>• <em>The Divine Comedy</em> by Dante Alighieri: <a href="https://www.amazon.com/dp/0451208633" rel="ugc noopener noreferrer" target="_blank">https://www.amazon.com/dp/0451208633</a></p><p>• AS Roma (football club Alessio and Claire are both fans of): <a href="https://www.asroma.com/en" rel="ugc noopener noreferrer" target="_blank">https://www.asroma.com/en</a></p><p>—</p><p><strong>Where to find Alessio Fanelli:</strong></p><p>X: <a href="https://x.com/FanaHOVA" rel="ugc noopener noreferrer" target="_blank">https://x.com/FanaHOVA</a></p><p>Latent Space podcast: <a href="https://www.latent.space/" rel="ugc noopener noreferrer" target="_blank">https://www.latent.space/</a></p><p>—</p><p><strong>Where to find Claire Vo:</strong></p><p>ChatPRD: <a href="https://www.chatprd.ai/" rel="ugc noopener noreferrer" target="_blank">https://www.chatprd.ai/</a></p><p>Website: <a href="https://clairevo.com/" rel="ugc noopener noreferrer" target="_blank">https://clairevo.com/</a></p><p>LinkedIn: <a href="https://www.linkedin.com/in/clairevo/" rel="ugc noopener noreferrer" target="_blank">https://www.linkedin.com/in/clairevo/</a></p><p>X: <a href="https://x.com/clairevo" rel="ugc noopener noreferrer" target="_blank">https://x.com/clairevo</a></p><p>—</p><p>Production and marketing by <a href="https://penname.co/" rel="ugc noopener noreferrer" target="_blank">https://penname.co/</a>. For inquiries about sponsoring the podcast, email [email protected].</p>
Key Insights
- Fanelli transitioned from being an 'agent prompter' to an 'agent manager,' which required moving agents from local machines to cloud VPS infrastructure with multiple communication channels (text, Linear, shell).
- Symphony is not providing new agent capabilities but rather a context-shaping framework—its primary value is organizing task history, specifications, and feedback in one searchable location compared to scattered conversations.
- Token usage tracking per task reveals inefficiencies: when actual token costs deviate significantly from expectations, the problem is usually in tooling or specification quality rather than agent capability.
- Models have a tendency to add rather than remove instructions in markdown files, causing specification bloat over time; Fanelli recommends regular purging of outdated guidance.
- For long-running AI tasks, response time is less valuable than 'clock time savings'—the ability for AI to autonomously complete work that would require human hours or days of manual effort.
- AI enables scaling of businesses based on heterogeneous data (trading cards, vintage clothing) that were previously impossible to automate with traditional classification or rule-based systems.
- Small specification files that define primitives and constraints work better than detailed prescriptive instructions; newer models naturally lock to the spec when executing work.
- Fanelli uses AI for stress reduction as a 'safety net' in personal finances and email management, offloading decision-making uncertainty rather than seeking optimization gains.
Topics
Transcript
This is my favorite positive outcome of AI, which is small business creation. Just the ability to like intersect the human world in a way that has been historically very inefficient has been a quality of life improvement for me. You know, my dad, their business, they deliver fish to restaurants. They got like this freezer with the frozen stuff and like somebody's going out there with like the pen and paper every morning, kind of like writing down what's there. Sometimes they're like, oh my God, we're missing like three tuna three tuna so like we're missing a box of shrimp all of that work now can easily be automated even with just with the metaclasses and you have…
Full transcript available for MurmurCast members
Sign Up to AccessMore from How I AI
This solo builder runs 24/7 local AI on his own hardware | Alex Finn
Alex Finn, a solo builder, discusses his setup of multiple local AI machines (Mac Studios, DGX Spark, RTX 5090) running 24/7 to power autonomous workflows for software development, security scanning, and market research. He explains why local models unlock unlimited AI usage compared to cloud APIs, and demonstrates his 'software factory' with automated build and review loops using Claude Code.
GPT-5.6 Sol vs. Claude Fable: Why OpenAI’s new model crushes my benchmark
A creator conducts a detailed benchmark comparing OpenAI's new GPT-5.6 models (Sol, Terra, Luna) against Claude Fable across real-world tasks like PRD writing, prototyping, and coding. GPT-5.6 Sol emerges as the superior choice for practical product development, offering better design uniqueness, clearer communication, and more effective collaboration despite Fable's theoretical intelligence.
What a harness is and how to build one with Claude Agent SDK
A harness is custom code built around an AI agent to make it more effective for specific workflows. The speaker demonstrates building a Sentry bug-fixing harness using Claude Agent SDK, explaining how structured constraints, opinionated tooling, and custom prompts enable AI agents to solve complex tasks more reliably than general-purpose coding tools.
Sonnet 5 review: I ran 64 generations to find out if it's worth it
The creator develops the 'How I AI Bench,' a custom evaluation framework combining human vibe checks with LLM judging to assess AI models on practical tasks like PRD writing, prototyping, and agentic coding. Testing five frontier models including Claude Sonnet 5, the results reveal significant disagreements between automated scoring and human preference, with Sonnet 4.6 and Gemini 3 Pro ranking highest despite Sonnet 5 being the newest release.