Key Takeaway
A meta-agent architecture driven by human-written ontologies can guarantee that global rules are never violated and that every state change is auditable, even if individual subagents are wrong or nondeterministic.
ON THIS PAGE
Key Findings
Design agents from a clear, human-authored ontology and give each subagent only a slice of state. Have a single meta-agent check proposed actions against declared global rules (invariants) before committing changes. That setup provably prevents coordination failures like double-spend or stale-state billing errors and produces an immutable, one-record-per-decision audit trail. audit trail. A human-in-the-loop queue can be added where regulations demand human judgment without breaking those guarantees. guarantees
Data Highlights
1Budget example: A $100,000 system-level budget is protected from conflicting agent actions (prevents overspend that would occur if two agents each spent $45,000 and $60,000).
2Testing limits example: In the billing case, patients in a 90+ day abstinence tier are eligible for 3 definitive tests per 90 days, while 0–30 day tier patients are eligible for 1 definitive test every 7 days—rules the meta-agent enforces to avoid billing mistakes.
3Audit guarantee: The audit log grows by exactly one immutable entry for every adjudication (approve, reject, or escalate), ensuring linear, tamper-evident records of decisions.
What This Means
Engineers building systems where multiple automated agents must coordinate and compliance teams in finance, healthcare, or security who need provable, line-by-line audit trails. Platform architects and operations leads can use this architecture to reduce runtime surprises and show regulators a formal safety argument. audit trails
Not sure where to start?Get personalized recommendations
Key Figures
Fig 1: Figure 1 : The Agentic Redux architecture. Sub-agents receive isolated state slices and propose actions to a meta-agent, which adjudicates proposals against declared invariants before committing state transitions. Code is available at https://github.com/Thistleseeds/agentic-redux
Fig 2: Figure 2 : Ontology-First Agent Design. The ontologies referenced in this paper were produced by this method. Ontology files are available at https://github.com/Thistleseeds/agentic-redux
Fig 3: Figure 3 : Role Derivation procedure.
Fig 4: Figure 4 : Deriving invariants, then assigning the cross-cutting invariants to the Agentic Redux meta-agent.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
Guarantees depend on a correct, human-authored ontology and initial state that satisfies the declared invariants; bad or incomplete ontologies weaken the protection. The work focuses on safety and auditability, not on system liveness (making progress) or performance overhead; a follow-up addresses liveness. Real-world deployments should still use defense-in-depth: wrap the architecture with runtime monitoring and human oversight to catch implementation bugs. runtime monitoring
Full Analysis
Start by defining the problem domain as an ontology (things, roles, and processes) written by a human expert. Derive agents that each hold a limited, well-defined slice of state and let them propose actions rather than directly mutate global state. A single meta-agent adjudicates proposals: it only commits transitions that preserve the domain's global invariants. That architecture—called Agentic Redux—prevents classic coordination failures like write skew (two agents independently making changes that together violate a global rule) and supports an optional human-in-the-loop queue for decisions that must remain human. coordination failures The guarantees are formal: nondeterministic runs (including unpredictable language model outputs) are modeled by an oracle trace and proven to preserve invariants for every reachable state. The system also provides audit log integrity: each adjudication produces exactly one immutable log entry recording the decision and rationale. Practical examples include a $100,000 budget scenario and detailed healthcare billing rules for urine drug tests. For real-world use, designers must invest effort in correct ontologies and should complement the architecture with runtime governance and monitoring; the approach is best suited to domains where provable safety and linear auditability matter most.
Explore evaluation patternsSee how to apply these findings
Credibility Assessment:
Single author with no affiliation or reputation signals, arXiv preprint — minimal identifiable credibility.