Controlled agents
Agent-style systems interleave reasoning with tool use. Making them safe for real operations is a permissions, validation, and audit problem, not a prompting problem.
- Operational use
- Multi-step operational work: gather, draft, check, file, where each consequential action is bounded, logged, and reversible or approved.
- When not to use it
- Open-ended autonomy over consequential systems. If an action cannot be bounded, validated, or undone, it does not belong to an agent.
- How it is validated
- Staged rollout with explicit permission tiers; recorded fixtures for the action layer; red-team style testing informed by the OWASP LLM Top 10; human confirmation for consequential actions.
- Limitations
- Risk frameworks are references, not certifications; Eigenvalue does not claim NIST compliance or OWASP certification. Prompt injection, excessive agency, and data exposure remain live risks that architecture, not vigilance, must contain.
ReAct established the now-standard pattern: a model that reasons, acts through tools, observes results, and continues. The pattern is powerful and, unbounded, unsuitable for operations that matter.
Eigenvalue builds agents inside an explicit control envelope: enumerated tools with least-privilege access; staged execution where consequential actions require confirmation; complete logs of reasoning, action, and result; and human takeover that halts automation cleanly. Design draws on the NIST AI Risk Management Framework and its Generative AI Profile for risk mapping, and on the OWASP LLM Top 10 for concrete failure modes: prompt injection, sensitive-data exposure, and excessive agency chief among them. These frameworks inform the architecture; they are not badges, and are not claimed as such.
Where it appears in our work
- Accounts Receivable and Collections Automation Relevant research reference
- Controlled AI Assistant and Engineering Foundry Implemented in this system
Primary sources
- Yao et al. (2022). ReAct: Synergizing Reasoning and Acting in Language Models
- NIST (2023). AI Risk Management Framework
- NIST (2024). AI RMF: Generative Artificial Intelligence Profile
- OWASP (2025). Top 10 for LLM Applications / GenAI
Last reviewed 2026-08-02. Sources link to primary or official pages; verify current versions before relying on them.