Agent retries that feed errors back into the prompt make every attempt cost more than the last

The retry state in the post's example is the conversation itself ^[3]. When the tool fails, the orchestrator feeds the stack trace back into the prompt and asks the model to correct its approach. By the third request the prompt holds the entire history ^[3]. The post traces the pattern to the default reaction of telling the system to "try again and fix the error" ^[12]. Each attempt re-sends everything before it plus the newest error, and cost per call rises with every failure ^[2].

The post says cost per call "snowballed exponentially" ^[5]. Its own figures trace a line. The first two retries add 600 and 700 tokens ^[1], and the climb from 2,000 tokens to 20,000 by attempt 50 averages about 367 tokens per retry ^[2]. When each call grows by a roughly fixed amount, the total across a loop grows with the square of the attempt count ^[5]. On the post's line, 50 attempts send about 550,000 input tokens, 275 times the first call ^[4][7]. Fifty retries at a flat 2,000 tokens would send 100,000 ^[8]. Appending the history multiplies the bill by 5.5 ^[9].

Leaving loop detection to the model ^[7] puts the stop decision in the one component that gets a longer error log each time it fails to make it. The post puts the observer outside instead: logic, execution and error handling split into distinct entities inside a strictly monitored state machine ^[11]. On placement, the post warns: "If the retry logic is housed inside the prompt itself, you have already lost control of the system." ^[9]

In my context I would set a per-task token budget first and a retry cap second. A retry cap bounds the attempt count, and under linear growth that also bounds the total ^[5]. A token budget also catches the single attempt that drags in an oversized page or trace. It counts the quantity the cost comes from ^[2]. Both have to count per task. A per-call ceiling set anywhere above 20,000 tokens would let the post's loop run all 50 attempts untouched ^[4].

For the post's token figures to carry over to another agent, its orchestrator has to append the full error output on every retry and keep the whole history ^[3]. An orchestrator that truncates or summarizes old errors grows more slowly per call. The post puts average first-year implementation costs for enterprise agentic AI at $45,000 to $250,000 and blames a large share on unchecked recursive loops ^[10]. It does not source that range, and its $542.18 weekend is written in the second person as an illustration ^[1].