Does a small team need all four controls?
Yes, but at the right scale. A small pilot can start with human review on every sensitive action and relax it as evals prove the agent is reliable.
Concept · Agentic security
The four production controls every AI agent needs
Before it reaches production, an AI agent needs four controls in place: reversible actions, human review for high-risk steps, searchable logs, and trusted evals.
Why production is not a demo
A demo runs in controlled conditions with a person watching. In production an agent acts on real data, connected tools, and decisions that affect customers. The risk surface grows, and failures stop being anecdotes — they become operational incidents.
- A demo proves something can work; production demands that it fail safely.
- Agents wired to email, CRM, or documents widen the attack surface.
- Without controls, a small error propagates before anyone sees it in time.
The four controls
Paput scores and keeps four controls in place before any agent reaches production. They are not optional — they are the difference between reliable automation and a fragile system.
- Granular rollback: every consequential action stays reversible, so a misbehaving agent can be undone before it reaches its limits.
- Human review queue: high-risk actions route to a named owner with clear exception handling, not an anonymous approval.
- Searchable logs: correlation IDs trace every decision across the chain, so you can reconstruct what happened and why.
- Trusted evals: quantified thresholds — an accuracy floor and an override-rate ceiling — catch silent drift before users do.
How they map to recognized standards
These controls are not invented from scratch: they operationalize frameworks security and compliance teams already use. That makes them auditable and defensible to third parties.
- NIST AI Risk Management Framework: risk across design, development, use, and evaluation.
- OWASP Top 10 for LLM Applications: prompt injection, sensitive-information disclosure, insecure output handling, excessive agency.
- CSA AI Controls Matrix: 243 controls across 18 domains.
Questions buyers ask
Do these controls slow the project down?
They add design work up front, but they avoid the far larger cost of an agent misbehaving in production with no way to undo it or know what it did.
AI operator field notes
illmethinks.io publishes source-transparent notes on AI agents, tools, and operational risk monitored by Paput.ai.