The Last Mile Is Code
summary
Agents reach coding-level performance (SWE-Bench climbs from 33 to mid‑80s, Mythos mid‑90s) so they can close the last mile by operating on infrastructure-as-code. Demo uses TypeScript, Pulumi-style IaC, S3 + CloudFront, Supabase and Edge Functions, and OpenAI for images; deployment finishes in minutes and images generate (sometimes slowly). Platform metrics show agent deployments rising 4% to 28% of all deployments. Concrete claim: make deployment steps scriptable so agents reliably ship TypeScript all the way down and run the full loop.
AI-written summary · about
chapters (11)
- 0:00Why Agents Love Code
- 1:32Why AI Excels at Software Engineering
- 3:10SWE-Bench and the Rapid Improvement of Coding Agents
- 4:35Does Code Still Matter?
- 5:27Vibe Coding a Trail App
- 6:45The Real Challenge: The Last Mile
- 8:13Infrastructure as Code for Agents
- 12:58Git Diff for Your Infrastructure
- 14:13CodeAct: Why Agents Should Write Code, Not Tool Calls
- 16:16Live Demo: Deploying an App with AI
- 23:13The Future of Agent-Driven Infrastructure
related talks
39m.NET: Infrastructure as Software
23mThe Agentic Infrastructure Gap: In-Distribution Languages Make It a Coding Problem
48mAI Coding Agents and How to Code Them
14mA Song of Types and Agents
30mBeyond the Hype: Using AI Coding Agents Today
36mDeveloper Experience Trends & Predictions
26mAI Agents as Your Co-Developer
19mBuilding Agents at Cloud Scale