Exacting A.I.: Harness Engineering for All

Moving from "AI-assisted Code Editor" to "AI-assisted Systems Architect"

This talk is an exploration on mitigating the human-review bottleneck, context rot & improving code correctness in production.

We explore what it means to adopt a 'systems architect' mindset for AI-assisted coding.

We’ve all experienced the magic of Cursor writing 500 lines of functional code in ten seconds. We have also all experienced the 3:00 AM terror of trying to debug a critical production failure inside a complex system we have no idea about.
As AI-assisted coding transitions from a neat party trick to the standard way we build software, the bottleneck has shifted.

The challenge is no longer generating code—it is in managing Cognitive Debt, context rot, and in managing our expectations 😅.

### About the Talk

In this session, we are looking past the "10x Developer" honeymoon phase to talk about Harness Engineering: the strict guardrails, workflows, and defensive architectures required to keep autonomous LLM agents from driving your codebase off a cliff.
Whether you are running multi-agent simultaneous coding, wrestling with Claude Code, or trying to scale a brownfield project inside Cursor, this talk will give you a framework for low-trust AI development.

Some snippets from the presentation:

• The "Process Smell" Test: Why not having to manually intervene during an AI generation wave is a big red flag even if your tests passed.
• Defensive Scaffolding: Implementing `.cursor/rules`, file deny-lists, and granular read-permissions to protect sensitive logic from rogue agents.
• Architecture-as-a-Leash: Using DSLs (fitness functions/Structurizer->Mermaid) to formulate a contract for the AI before it is allowed to write the code and to map "Current-State" / "Target" diagrams*.*
• Taming Multi-Agent Fan-Out: How to manage context rot and bleed by enabling context isolation with gsd-core that safely routes agent(s) through Research → Planning → Execution → Verification
• Toolchain reality check: Pragmatic tips for coding workflows, and GSD-core orchestration.
• Increasing your information bandwidth: How to keep up with thousands of lines generated by your Junior AI developer.

Who this is for: Software Engineers, students and Indie Builders who use Cursor, Copilot, or Claude daily, who have already blown past the "hello world" stage of AI coding, and have felt the gradual, heavy friction of hitting the complexity ceiling due to production bugs and human review bottlenecks.