Brownfield Engineering — Shipping AI into Legacy Codebases | The Stage @ SQ Collective
Hrishi, CEO of Southbridge.ai, on why brownfield engineering — old code, tangled state, real users — is the hardest place to ship reliable AI agents. Demos are a souffle for friends. Brownfield is a professional kitchen.
Hrishi covers what the model can actually see (the ~1MB useful-context ceiling), why frontier models hold 400K+ tokens on a single unfolding problem but fail at much smaller "haystack" tests, and a CIA-sabotage-manual anti-guide to keeping your AI project stuck in greenfield forever (never delete from context, rebuild from scratch, pass knowledge model-to-model). For engineers shipping AI into production code.
Recorded live at SQ Collective on 2026-05-01.
Chapters:
00:00 Intro
00:19 Background: Southbridge & Data Agents
02:19 Problem: AI Context Windows
03:09 Bottom of the Stack AI
05:58 How NOT to Do AI: Huge Context
06:48 How NOT to Do AI: Greenfield Only
07:38 How NOT to Do AI: Vibe Coding
08:48 How NOT to Do AI: No Abstraction
09:38 How NOT to Do AI: Working in Silos
10:09 Hankweave: Abstracting AI Programs
13:14 Q&A: Team Coordination
13:46 Q&A: Collecting Failure Projects
15:40 Q&A: Generalizing Coding Agents
18:09 Q&A: The Pace of AI Improvement
20:20 Q&A: True Context Limits
21:55 Q&A: Hank Runtime & Sentinels
27:54 Q&A: Open Source Models
29:14 Q&A: Sentinel Evaluation Loops
Full recap: https://www.sq-collective.com/post/wh...
SQ Collective is Singapore's coworking space and community for AI builders, founders, and product teams.
→ Next event: https://lu.ma/sqcollective
→ Site: https://sq-collective.com
#SQCollective #AIAgents #ContextEngineering