Start with the session evidence before changing budgets or models. Identify whether the increase came from retries, idle activity, longer sessions, or more users, then test the most likely control failure. If retries dominate, authentication and connector stability deserve attention before broader optimisation work.
Start with the evidence that explains the cost spike
When agent costs rise unexpectedly, the first job is to separate usage growth from control failure. Look at session evidence, not just invoices: retries, idle activity, longer sessions, and more users produce very different remedies. That distinction matters because cost spikes often reflect a behavioural or integration problem before they reflect a pricing problem.
The practical test is to identify which pattern changed first. If retries dominate, inspect authentication failures, connector instability, and repeated tool calls before you broaden the optimisation effort. If idle time or session length dominates, focus on orchestration, timeout policy, and user workflow design rather than model selection alone.
Teams should also verify whether the increase is concentrated in one workflow, one connector, or one user population. A narrow spike usually points to a broken dependency or a misconfigured control; a broad spike is more often a genuine workload shift or a governance issue around agent usage.
Why retry-heavy spend usually points to a control problem
Retry-heavy spend is rarely a pure efficiency issue. It often means the agent is repeatedly failing to authenticate, reach a tool, complete a step, or preserve context, which turns every failed attempt into extra spend and more surface area for error. In that case, cost is a symptom of reliability and access friction, not simply model choice.
This is where the session trail is more useful than aggregate token or request counts. If the same request pattern is being repeated, the control failure is usually closer to the access path than the model itself. That can include unstable connectors, expired sessions, broken tokens, weak handoff logic, or repeated human corrections that cause the agent to re-run work.
For teams running agentic systems, this is a good moment to use a structured security view such as Agentic AI Security Guide and AI Agent Observability, Audit and Incident Response Guide, because both help separate noisy spend from genuine control failure.
What to do before you optimise models or budgets
Start by establishing a baseline for normal session duration, retry rate, and tool-call volume, then compare the spike against that baseline. That gives you a defensible answer to whether the issue is demand, reliability, or policy. Once you know that, the response becomes more targeted and less expensive.
Where the spike is tied to authentication or connector instability, the first fix should be the broken control path, not the budget. In practice that means checking whether the agent is re-authenticating too often, whether a connector is failing open or failing closed in a costly loop, and whether session handling is causing repeated work. For environments using delegated access, AI Agent Authorisation Guide and Zero Trust for AI Agents are useful references for tightening access per action and removing standing privilege.
If the spike is caused by broader usage growth, then optimisation can come later, after you have confirmed the demand is legitimate. That sequencing avoids spending time tuning models to compensate for an operational fault that would keep driving cost even after the tuning work is done.
Risk and Threat Considerations
Unexpected agent spend is not only a finance issue, it can be a sign of compromised, misrouted, or overly permissive agent behaviour. Repeated retries, long-lived sessions, and connector instability can hide abuse, inflate operating cost, and make it harder to tell normal workload growth from malicious or accidental overuse.
Failure mechanism: An agent that repeatedly re-authenticates, replays tool calls, or loops through a failing connector can generate a compounding cost pattern while masking the original fault.
Impact: Teams can miss a real access or reliability problem, overpay for an avoidable workload pattern, and leave the same failure path available for wider abuse.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Agentic AI Top 10 addresses the attack and risk surface, while NIST SP 800-53 Rev 5 sets the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Agentic AI Top 10 | ASI02 — Tool Misuse | Retries and broken tool loops can drive agent spend and control failure. |
| ASI03 — Identity & Privilege Abuse | Authentication and access instability can surface as retry-heavy spend. | |
| Recommendation — Inspect tool-call loops and add guardrails that stop repeated failed actions. Tighten per-action authorization and eliminate standing agent privilege. | ||
| NIST SP 800-53 Rev 5 | AU-6 — Audit Record Review, Analysis, and Reporting | Session evidence is needed to distinguish retries, idle activity, and growth. |
| IA-5 — Authenticator Management | Authentication problems are a likely source of repeated retries and cost growth. | |
| AC-6 — Least Privilege | Overly broad access can increase unnecessary agent activity and spend. | |
| Recommendation — Review session logs to isolate the first control failure behind the cost spike. Stabilize authenticator lifecycle and renewal behaviour before broad optimisation. Reduce standing access so agents can only invoke the actions they need. | ||
Practitioner Guidance
What to verify: Confirm whether the spike is driven by retries, idle sessions, longer sessions, or higher user volume before changing model settings or budgets. If retries dominate, inspect authentication health, connector stability, and session renewal behaviour first.
Decision rule: If the cost increase is concentrated in one workflow or connector, treat it as a control-failure investigation. If it is broad across users and sessions, treat it as a workload-change problem and recalibrate capacity and policy together.
Practitioner takeaway: The fastest way to avoid waste is to diagnose the session pattern first, because cost spikes caused by retries or broken access controls will not be fixed by generic optimisation.