Insights & Updates

Thoughts on AI agents, capability systems, structural safety, and the future of software.

ResearchJune 2026 · 18 min read

Your AI coding agent ships debt 69% of the time when you're not looking

We ran coding agents in unsupervised edit-until-done loops and scored the final repository against a deterministic oracle. Unguarded loops shipped net-new debt in 11 of 16 runs; a prompt-only self-check reached only 9 of 16; an external deterministic gate eliminated observed escapes (0 of 16). The binding constraint is specification and verification, not model capability.

Read
By Sid
ArchitectureJune 2026 · 20 min read

Where the Proof Comes From: The Operating Graph Behind Proof-Carrying Operations

Why trusted AI action requires build-time structure: resolved identity, cited rules, bounded capabilities, evidence packets, and deterministic admission. Part two of the proof-carrying operations series.

Read
By Sid
ArchitectureJune 2026 · 16 min read

Proof-Carrying Operations: From Digital Twins to Trusted Action

Digital twins are necessary but not sufficient for industrial AI agents. Proof-carrying operations is an architecture where every high-stakes AI-assisted action carries evidence, rules, approvals, and an audit trail before it crosses into operation.

Read
By Sid
Case StudyMarch 2026 · 11 min read

SAP Clean Core Migration: Ontology Extraction in Practice

We connected Vyuh to a live SAP system and extracted its full ontology: custom objects, Fiori apps, OData services, data flows. Then we deployed an AI advisor grounded in the extracted structure. Here's what we learned.

Read
By Sid
ArchitectureFebruary 2026 · 14 min read

Why Agent Guardrails Don't Work (And What Does)

Everyone building AI agents hits the same wall: how do you let an agent act without it doing something catastrophic? Guardrails feel right. They're also fundamentally flawed.

Read
By Sid
IndustryFebruary 2026 · 12 min read

From Chatbots to Compilers: The Evolution of AI Agents

We kept feeding models more leverage. More context, more tools, more autonomy. Until the bottleneck moved from model intelligence to everything around the model.

Read
By Sid
TechnicalJanuary 2026 · 10 min read

Code Understanding is Harder Than Code Generation

Everyone's excited about LLMs writing code. But the harder problem is understanding code that already exists. Generation is greenfield. Understanding is brownfield.

Read
By Sid