RIZZ: Routing Interactions to Near Zero-Interference Zones for Continual Adaptation of Black-Box Agents

Researchers introduce RIZZ, a black-box adaptation framework for large language models deployed as long-lived agents that must continually adapt across diverse tasks and domains without access to model weights. The system uses verifier-gated memory, dynamic routing, and prompt compilation to prevent task interference while learning from sparse feedback in nonstationary environments.

RIZZ addresses a critical challenge in deploying large language models as autonomous agents: maintaining performance across diverse, shifting contexts without the ability to fine-tune or modify the underlying model. This research tackles the tension between continual learning and catastrophic forgetting—when an agent learns new tasks but loses proficiency in previous ones. The framework's innovation lies in its architectural approach: dynamically spawned memory branches isolate task-specific learning while context-aware routers prevent knowledge contamination between unrelated domains.

The significance stems from the growing deployment of LLMs in production environments where users, tasks, and feedback regimes vary unpredictably. Current approaches either optimize single prompts statically, maintain undifferentiated memory that causes interference, or rely on computationally expensive rollout-heavy search methods. RIZZ's verifier-gated system ensures only validated interactions update memory, preventing harmful patterns from propagating. This is particularly valuable in resource-constrained settings where online adaptation must occur within strict context budgets.

For developers and enterprises deploying LLM agents, this represents a meaningful advancement in reliability and efficiency. The ability to serve diverse user bases without task interference reduces failure modes and improves agent robustness. The black-box approach—requiring no access to model weights—makes it immediately applicable to proprietary models and API-based deployments, expanding its practical utility across the AI industry.

Key next steps include evaluating RIZZ performance on long-horizon, multi-domain benchmarks and examining computational overhead of the routing and memory management systems. The framework's success could influence how production LLM systems are architected for continual learning applications.

• →RIZZ enables black-box LLM agents to adapt continuously across tasks without access to model weights or catastrophic forgetting.
• →Dynamic memory branching with context-aware routing isolates task-specific knowledge while preventing cross-task interference.
• →Verifier-gated memory updates ensure only validated interactions improve agent behavior, reducing failure propagation.
• →The framework targets resource-constrained deployment scenarios requiring online adaptation within bounded context windows.
• →Architecture is immediately applicable to proprietary and API-based LLM deployments in production environments.