[Feature]: Add Economic Safety Guardrails to Prevent Token / Cost Exhaustion (Denial of Wallet)Β #27
Description
Feature Request
Introduce economic safety guardrails as first class citizens in CUGA, such as:
- Token / cost budgets per task or session
- Output size or precision limits (policy-driven)
- Preflight estimation of large responses
- Graceful degradation instead of runtime failure
Motivation / Problem
CUGA can be driven into excessive token generation by non-malicious requests that gradually increase output size (e.g. higher numeric precision).
The agent repeatedly complies, expands the response, and eventually fails with a context-length error β after incurring significant token cost.
This is a classic Denial-of-Wallet / Economic Abuse scenario
See linkedin post from Divyanshu Kumar https://www.linkedin.com/posts/divyanshuusingh_aiagents-enterpriseai-cybersecurity-activity-7407839489756205056-k7np
Use Case
"...When I finally asked for 1,000,000 decimal places of Ο, the agent again attempted to comply, aggressively expanded the response, exhausted the context window, and failed with a context length exceeded error.
Nothing malicious. Just expensive...."
Proposed Solution
The agent should proactively detect and constrain tasks that imply unbounded or escalating output, and respond gracefully (cap output, renegotiate, summarize, or refuse with explanation) instead of failing at context overflow.
Alternatives Considered
No response
Priority
Medium - Would be helpful
Implementation Complexity (if known)
None
Additional Context
No response
Checklist
- I have searched existing issues and feature requests to ensure this is not a duplicate
- I have provided a clear use case and motivation for this feature
- I am willing to help test this feature once implemented
- I am interested in contributing to the implementation of this feature