We Built This Before O’Reilly Said It Was a Problem

On March 26, 2026, five speakers at O’Reilly AI Codecon independently described the same unresolved problem in AI agent systems. From five different angles — reliability engineering, cognitive science, agent provenance, production failure, and capability architecture — they each arrived at the same gap.

We’d already built the answer. Here are the receipts.

“The problem shifts from ‘can the model solve this?’ to ‘can the system execute a stable sequence of decisions?’”

Nicole Koenigstein, March 26, 2026.

She introduced Lusser’s Law applied to multi-agent systems: chain 10 agents at 98% individual accuracy and your system accuracy drops to 81.7%. The hidden cost isn’t just compute — it’s the compounding of silent failures across a pipeline nobody is monitoring at the decision layer.

We published our six-gate constitutional architecture in January 2026. Gate 4 (EPG) evaluates economic sustainability of every decision chain. Gate 1 (EG) catches epistemic overconfidence at each step. The gates don’t just block bad outputs — they interrupt unstable decision sequences before they compound.

Nicole named the problem. Commit d5071bfb (February 2026) was us solving it.

“Without identity and lineage, an agent is not reproducible. And if it’s not reproducible — it’s not governable.”

Tatiana Botskina, March 26, 2026.

She drew the analogy precisely: distributed systems needed DNS (identity), TLS (trust), and Git (lineage) before they became governable at scale. AI agents have none of this. We’re running autonomous actors with no system of record.

Our constitutional framework mandates what Tatiana describes: every agent has a declared capability manifest, authority level, and audit trail. Amendment 49 (December 2025) ratified sub-agent governance with explicit tiers. The NIST AI Agent Identity concept paper we submitted in March 2026 makes this case formally.

Tatiana’s line belongs in every AI governance conversation. We’ve been building the infrastructure she’s describing for three months.

“Instrument use is rare — only 5 to 10 percent of AI interactions.”

Mike Amundsen, March 26, 2026.

He mapped three modes of AI use: Generator (60–70%, get answers), Surrogate (20–30%, get advice), Instrument (5–10%, support and challenge your own thinking). Most AI use is Generator mode. Judgment shifts from humans to machines. Skill formation weakens.

This is the cognitive cost we measure with the Decision Load Index. The DLI doesn’t ask “did AI help you?” It asks “what did AI oversight cost your brain?” Generator mode users carry the highest cognitive overhead — they’re reviewing and validating AI output without developing judgment. That’s the problem BCG named “AI Brain Fry” in March 2026 (HBR, Fortune, CNN coverage).

The DLI is the measurement instrument for the gap Amundsen described. Published: DOI 10.5281/zenodo.18217577.

“The failure wasn’t catastrophic. It was insidious.”

Advait Patel, March 26, 2026.

Broadcom deployed background AI agents that analyzed logs and proposed fixes. One agent generated a change that passed all tests, merged cleanly, and looked correct — then caused subtle production degradation hours later. Not a crash. A drift.

Our Ralph Loop resilience protocol (Amendment 8.6.7, January 2026) was built for exactly this. External verification hooks, behavioral sign markers, circuit breakers that open on pattern deviation rather than outright failure. The gates don’t just catch crashes — they catch drift. The distinction matters more than it sounds.

“Skills fossilize when AI handles them.”

Juliette van der Laarse, March 26, 2026.

Her AI Flower framework maps 145 engineering activities for AI-native organizations and includes a Skill Fossilization Model: when AI handles a task repeatedly, the human skill for that task atrophies. The capability architecture has to account for what’s being lost, not just what’s being gained.

This is the other side of the DLI. The DLI measures cognitive overhead — the cost of AI oversight. Skill fossilization measures cognitive atrophy — the cost of AI delegation. Two directions of the same phenomenon. A complete cognitive cost model needs both.

We’re building the atrophy measurement layer next.

What This Means

Five independent researchers, at a major O’Reilly conference, described five failure modes in AI agent systems. Each maps to a design decision we made in production months earlier:

We didn’t build this because O’Reilly validated it. O’Reilly validated it because the problems are real.

The governance gap in AI agent systems isn’t theoretical. It’s running in production, compounding silently, and costing organizations in ways they’re not measuring.

We built the measurement instrument. We built the governance architecture. Both are open and published.