Design a Cost and Loop Watchdog for Autonomous Runs

ROLE Act as an agent runtime reliability engineer. OBJECTIVE Design a resource watchdog for [AGENT NAME]. CONTEXT Allowed actions and tools: [TOOL INVENTORY]. Normal run examples: [RUN EXAMPLES]. Spending and runtime limits: [LIMITS]. Available telemetry: [TELEMETRY]. INSTRUCTIONS Define per-run and aggregate budgets for model calls, tokens, tool invocations, elapsed time, and paid external actions. Detect repeated arguments, cycling states, repeated retrieval without new evidence, and outputs that do not advance the task. Distinguish useful iteration from nonprogress using observable signals. Specify warning, graceful-stop, and hard-stop behavior, including what intermediate work to preserve. Explain how shared budgets work across parallel branches and how delayed usage reporting affects enforcement. Use supplied prices only; otherwise provide formulas. OUTPUT FORMAT Deliver a watchdog specification, threshold configuration template, pseudocode, and five synthetic run traces with expected interventions. QUALITY CHECK Flag assumptions and avoid fabricated statistics. Label suggested thresholds as provisional. Confirm that stopping a run does not trigger automatic retries that recreate the same spending loop.