Control · C-106 · Security & Resilience
Cumulative limits on agent actions
Each agent works within limits on the total count, value and rate of its actions over a period, enforced outside the agent, with a hard stop when a limit is reached.
- Control type
- Preventive
- Family
- Risks addressed
- 5
- Audience
- Heads-of · Practitioner
- Frameworks
- ASI02 · ASI08
- Responsible AI domain
- Lifecycle stage
- Reviewed
- 2026-11-17
What it does
It bounds the combined effect of many actions, as approval gates bound the effect of one. However an agent is misled or malfunctions, the total it can do in an hour, a day or for one customer is capped. The limit sits in the component that carries out the action and not in the agent, so the agent cannot reason or be talked around it. The business owner sets the limits as they would set a delegated authority for staff; engineering enforces them.
How to implement it
- For each consequential operation, set limits on count, value and rate per period, and per customer or target where relevant.
- Enforce them in the gateway or service that carries out the action, outside the agent.
- At a limit, stop the agent and route further actions to a person; do not merely alert.
- Count retries, sub-agents and parallel instances towards one total.
- Do not reveal limits to the agent or to users; a known threshold becomes a target.
- Review limits when volumes or the agent's scope change.
Evidence it operates
Design: a register of limits for each agent and operation, approved by the business owner, stating where each is enforced. Operation: logs of limits reached and of the agent stopping; a test in which an agent is driven past a limit through many small actions and through parallel instances, and is stopped both times.
Where it fits
Set at design and adjusted in operation. It is the primary control for harm from cumulative agent actions. It complements approval gates for high-impact agent actions, which deal with single actions above a threshold, and inference usage quotas and rate limiting, which limit use of the model and not the effect of actions. It relies on the identifier carried by agent action tracing to count the actions of sub-agents and parallel instances as one total.
Limits
A limit set too high gives little protection and one set too low stops legitimate work, so the setting needs data. Within the limit it cannot tell a harmful pattern from a busy day. Totals must be counted across every route an action can take, which is hard where agents act through several systems.
Risks this mitigates
Primary mappings
- Harm from cumulative agent actions High
- Caps count, value and rate over a period, outside the agent. An agent takes many actions that are each within its limits, and their combined effect is one that no approver would have allowed.
Supporting mappings
- Agent tool misuse High
- Caps how much a misused tool can do over a period. An agent uses a tool it is permitted to use in an unsafe or unintended way, such as a destructive command or a wrong parameter, causing data loss, unauthorised transactions or disruption.
- Cascading failures in agent workflows High
- Stops a loop or cascade before its total becomes material. An error or compromise in one agent or step propagates through chained agents and automated workflows faster than people can notice or intervene.
- Excessive agency in autonomous agents High
- Caps the total effect of whatever authority the agent holds. An agent is given more functions, permissions or independence than its task requires, so an error, an ambiguous instruction or a manipulated input turns into a damaging action that no one approved.
- Unauthorised commitments by agents High
- Caps the value an agent can commit over a period. An agent makes a statement, offer or transaction that binds the organisation, without the authority a person in the same position would need.
Framework and clause references
| Framework | Clause | Title |
|---|---|---|
| OWASP Top 10 for Agentic Applications | ASI02 | Tool Misuse and Exploitation Bounds the total that misused tools can achieve. |
| OWASP Top 10 for Agentic Applications | ASI08 | Cascading Failures Stops a cascade at a limit. |
Related terms
How to cite
Axiom Verity (2026). Cumulative limits on agent actions. Reference Library, C-106. https://axiom-verity.com/controls/cumulative-limits-on-agent-actions. Reviewed 2026-11-17.
Found an error or a better source? Report a correction.