How Droid Handles 7M Tokens | Luke @Factory

Description

Luke Alvoeiro from Factory AI explains how Droid's agent scaffolding enables 7M token sessions without context resets. Technical deep dive into the architecture behind what I've been using. After my post about switching to Droid hit 70K+ views, I wanted to understand the technical "why." So I brought in Luke - he built Goose at Block (used by 5,000 engineers) and now architects Droid's context system. This is me learning the architecture alongside you. Full transparency: Factory AI is sponsoring this conversation. I asked for it because I genuinely want to understand what's happening under the hood when I run these extended sessions. Topics: Incremental summarization, agent scaffolding, context compression, parallelization, and why this matters for practitioners shipping real code. Try Droid with 40M tokens: https://rfer.me/rayfactory (Extra allocation specifically for my audience to test long sessions) Tired of AI giving you garbage code? I'll be your senior dev for 5 days. → The AI Architect Intensive: https://rfer.me/1337 What you get: ✅ 5 days of 1-on-1 senior guidance (same time daily) ✅ Fix your broken prompting permanently ✅ Ship real features by Day 5 ✅ Your money back if you don't ship working code Only $1,337 • 3 spots per week • US/Europe/Asia welcome SUPPORT THE CHANNEL Join 220+ builders in our private Discord where we ship AI apps together: https://rfer.me/discord This stream is powered by Ray Transcribes. Get professional, AI-powered transcripts for your content. https://raytranscribes.com Wispr Flow (Voice to Text): https://rfer.me/wispr 🕒 Key moments: 00:00:00 Why Droid for long coding sessions 00:00:41 The 7 million token session problem 00:02:11 How Droid avoids context bottlenecks 00:03:24 Droid's context compression strategy 00:05:52 What persists after context compression 00:07:57 How Droid respects Agent.MD files 00:11:28 Spec Mode vs. other agents' Plan Mode 00:16:51 Luke's background building 'Goose' at Block 00:23:49 Technical deep dive on anchored summaries 00:27:44 Explaining 'Agent Scaffolding' 00:30:29 How model-specific tools improve performance 00:38:00 The challenges of agent parallelization 00:46:53 Q&A: Using static call graphs 00:48:08 Q&A: Tips for token-efficient planning 00:49:52 Q&A: When to start a new session 00:52:01 Q&A: Is a Droid SDK coming? 00:57:41 Factory's enterprise business model 01:00:14 The future: A decade of agents 01:03:51 How AI will change dev practices 01:11:46 Building custom models vs. superior tooling 01:16:24 Luke's advice for junior engineers WHAT YOU'LL LEARN: - Luke's background: From Goose at Block to Droid at Factory - The context window bottleneck every practitioner hits - How incremental summarization actually works - Agent scaffolding: What it means architecturally - Anchored summaries vs full re-compression - T_max and T_retained thresholds explained - What gets preserved vs compressed in long sessions - Parallelization in agent workflows - Spec mode: Different approach to agent planning - Real examples from my 7M token sessions - Production considerations: cost, latency, quality - Your technical questions answered by the developer TOPICS COVERED: ✓ Goose at Block: Lessons from 5,000 engineer deployment ✓ Why context management is the real bottleneck ✓ Technical breakdown: How Droid's approach differs ✓ Incremental vs full re-compression explained ✓ What happens at the compression threshold ✓ How the system decides what to preserve ✓ Agent scaffolding beyond prompting ✓ Parallelization: When agents run concurrently ✓ Spec vs Plan architectural philosophies ✓ Enterprise integrations beyond code ✓ Where agent architectures are heading GUEST: Luke Alvoeiro - Developer AI @ Factory (ex-Block) - Built Goose: Open-source agent used by 5K Block engineers weekly - 3.5 years at Block shaping AI developer strategy - Now architecting Droid's agent scaffolding at Factory AI - Expert in context management and production agen