Investigating how AI agents
should be built and governed.
Mode's research program addresses open empirical questions in AI agent safety, orchestration enforcement, proactive behavior, context management, and permission systems. Each research topic is directly connected to how Mode Agent is built and where the field is going.
Research topics
Six active research areas. Each addresses a gap in the existing literature that Mode is uniquely positioned to study empirically.
Orchestration-layer enforcement vs. prompt-based approaches
Does enforcing permission boundaries at the agent orchestration layer reduce constraint violations more effectively than prompt-based approaches in multi-step autonomous AI workflows, particularly under adversarial prompt injection conditions?
The gap. Runtime enforcement has been studied in single-agent, single-model settings. The multi-agent, multi-provider orchestration case is unaddressed.
Human oversight calibration in autonomous multi-agent workflows
What frequency and granularity of human review checkpoints preserves acceptable safety properties while maximizing autonomous throughput, and can that Pareto frontier be empirically derived as a reusable benchmark?
The gap. Autonomy levels are defined categorically. The safety thresholds that justify moving between them have not been empirically validated.
User-programmable behavioral rules at the harness layer
Do user-defined behavioral rules enforced at the harness layer, below the context window, above the orchestration floor, produce more persistent and reliable compliance than equivalent rules expressed through prompts or memory systems?
The gap. Existing runtime enforcement is developer-facing. User-programmable rules below the context window have never been studied or benchmarked.
Orchestration-driven proactive agent behavior
Can orchestration-layer feedback loops enable AI agents to surface relevant information, initiate appropriate actions, and exhibit improvement-oriented behavior without explicit user prompting or model retraining?
The gap. Proactive AI research addresses model-level changes. Orchestration-layer approaches to proactivity without model retraining are unexplored.
Distributed context management across heterogeneous models
How does distributing context across heterogeneous models in multi-agent workflows compare to single-model full-context approaches in task completion, semantic fidelity, inference cost, and safety property preservation?
The gap. Context research is almost entirely single-model. No empirical benchmark compares distributed multi-model context against single-model alternatives.
Minimum viable permissions for AI agent task classes
Can task-type classification automatically derive minimum permission sets for AI agents operating across a multi-dimensional permission system, and does least-privilege enforcement reduce constraint violations without degrading task completion?
The gap. Least-privilege is foundational in OS security. Its application to multi-dimensional AI agent permission systems has not been studied.
Why Mode does this research
Each research topic connects directly to how the Agent is built. The Agent routes tasks across 600+ models from 100+ providers, enforces permission boundaries at the orchestration layer, runs multiple specialized agents in parallel, and manages context across model boundaries. The questions Mode investigates are questions Mode has to answer to build the product, and the answers don't exist yet in the published literature.
Published findings, methodology, and benchmark data are made available to the research community regardless of Mode's commercial trajectory. The system being measured is the one being built, but the results belong to the field.