agentsop-bounded-loop
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill consists of best-practice documentation for improving the reliability and security of agentic workflows. It provides specific implementation patterns to prevent billing runaway and infinite loops by ensuring termination logic is controlled by the framework rather than the language model.\n- [INDIRECT_PROMPT_INJECTION]: The skill describes workflows that process potentially untrusted external inputs such as tool errors or multi-agent handoffs. It provides mitigations by requiring that loop termination logic reside outside of the language model's control, preventing adversarial data from triggering infinite cycles.\n
- Ingestion points: Shared state fields like
messagesandlast_errordescribed inSKILL.md(Step 4).\n - Boundary markers: Recommends structured prompt formatting with explicit iteration counts and error context (e.g., 'Attempt {n} of {N}').\n
- Capability inventory: Tool invocation and multi-agent delegation capabilities inherent to the ReAct and planning loops discussed.\n
- Sanitization: Enforces termination through deterministic Python-based predicates and progress witnesses (e.g.,
dict.get('retries') < N) rather than relying on LLM-driven self-reflection.
Audit Metadata