Stop Runaway Agents before they turn into unexpected costs
SAVI watches each agent's call rate, output size and spend as calls arrive, flags a loop pattern, and sends a plain-language alert with the cost context. This page covers runaway agents. For all spend types, see Anomaly Detection.
- Detects looping behaviour as calls arrive, against each agent's own history
- Cost and reliability scores show what kind of problem you have
- Every incident explains which signals fired and why
Autonomous agents can quietly use up your AI budget overnight
Modern AI agents are autonomous: they retrieve documents, call APIs, summarise results and loop until done. Without guardrails, one misconfigured agent can make thousands of identical tool calls through the night. Finance sees the damage on the monthly invoice, not at 11:43 PM when it started.
One runaway loop adds up fast
A loop that runs all night can produce a very large bill. Standard budget alerts fire daily or weekly, after the damage has built.
Triage by hand takes hours
Without runtime detection, engineers read logs, find the looping agent and redeploy, which costs hours on top of the spend.
Post-mortems need a record
Without per-call logging you cannot show what happened, when it started or what it cost. A record makes the post-mortem quick and factual.
Detection pipeline
Every call checked against the agent's own history
SAVI scores each AI request as it arrives. It compares an agent's call rate with its own history and watches for output far larger than the input. Signals are combined, and a critical alert needs more than one of them to agree. Different workloads behave differently, so SAVI uses thresholds set for each type. More signals, such as repeated tool use, are being tested.
- A small SDK wrapper for the major providers
- Thresholds set for each workload type
- A clear path from normal activity to detection, alert and investigation
Flag, alert and investigate without stopping agents
When a threshold is crossed, SAVI flags the incident and sends a structured alert. Pausing an agent is an opt-in, so you decide how much SAVI does on its own.
- Alerts to Slack, PagerDuty, email or a webhook
- A live view of incidents as they are detected
- Alert context includes the signals that fired, the score and the cost so far
Example alert in a finance channel
Swipe to see all columns →
Example data. Agent loop events, 5 of 23 this month
A tamper-evident record of every detection
SAVI records every detection event: the agent, token counts, cost impact and the response. Entries in the audit log are signed, so an edit after the fact shows. When audit or the board asks what your AI did and what it cost, the record is already there.
- Signed entries, so changes after the fact show
- Export as JSON, or send to your security tools
- Role-based access, so people see what their role needs
Plain-language explanations on every incident
Every alert comes with a plain-English explanation of what triggered it, written for operators and not only engineers. SAVI tells you which pattern fired, how far above normal it was, and what it means for your AI budget. You do not need to decode scores or read log files.
- Every alert names the pattern that fired, how far above normal it was, and the estimated cost impact
- Plain-English summaries for operators and compliance reviewers
- Mark an alert as a true or false positive to help tune detection
Example alert detail
What detection gives your team
Do you know if your agents are using your AI budget well right now?
6 questions. Answer honestly.
Can you see every LLM call across your agents as it happens, not on a daily report?
Do you know which agents call the same tool with identical inputs again and again?
If an agent started a loop at 11 PM tonight, would you know before 9 AM tomorrow?
Can you spot a runaway agent without waiting for an engineer to go on call?
Do you have an audit trail of AI agent actions that your compliance team can open today?
Could your organisation absorb a large runaway loop that was found days after it started?
This checks one slice of your AI estate. Take the full AI Cost Ownership Assessment →
Your AI gap score
Answer to begin
Complete all questions to see your result
Find runaway agents when they start, not on the invoice.
SAVI detects looping agents as calls arrive and sends plain-language incident reports your compliance team can use.