An AI agent writing code is a persuasive reviewer of its own work. It can explain why a change is correct with great fluency, and a fluent explanation is not evidence. As AI agents take on more engineering and research, independent verification becomes the limiting factor, and NOVAQ decided to build that capability as a system of its own.
The design
Eight roles work as a team: orchestration, red team, strategy, account rules, simulation, execution, development and quality control. The red team exists to attack conclusions and reproduce them from scratch. Work moves through file-based tickets and an append-only ledger, so the history of each decision can be read later.
Three rules shape everything. A result without a manifest does not exist: a run has to record its inputs with hashes, its code, its environment and its outputs, and an unrecorded result cannot be cited. Each conclusion needs a second signature from another agent. And the swarm is read-only by default, with any action on a live system requiring explicit operator approval. Fixes go through a separate development and review cycle, and they need their own sign-off.
Eight roles around one recorded ledger
The scale
Since 5 September the swarm has completed 758 runs, 311 verification files and about 100 numbered findings. A change-history forensic pass inventoried 1,707 relevant commits and cross-linked 598 defect records from seven sources, with completeness tested by a canary and two-sided sampling. That gives NOVAQ a connected view of how the system changed and why.
What it caught
The most valuable finds were failures that raise no error, the kind that ordinary tests and monitoring rarely show. They included a family of time-zone problems that shifted one anchor by hours, skipped signals that left no trace across 18 call sites, and frozen data paths still being read. Each of these looks like normal operation from the outside.
It checks itself too
The swarm discovered that variants of one signal had been counted as several trades. Once corrected, a strategy family that had looked worse no longer showed a robust difference from its backtest. An audit system that can revise its own conclusions earns more trust than one that only ever adds findings.
From audit to engineering
In October the swarm gained development and quality-control roles, which made it a full improvement loop. The execution speed-up came out of that loop, with each batch checked for identical outputs. The same reproducibility rules support exact trade reproduction.
Why it matters
The audit swarm shows that multi-agent AI can do more than generate code. With clear roles, recorded evidence and separate sign-off, it can carry out sustained, reviewable verification on a real system. The design applies to any complex software where errors stay silent.
Limits
AI auditors can share blind spots with AI builders, especially when one vendor's model checks its own output. The human operator remains the final authority.