Three workflow bugs made one team's GPT-5.4 agent look lazy in production

The three faults stack. The retry cap fell by two-thirds ^[1]. The post argues that two attempts is often a trap for research, debugging or any tool use. One bad retrieval plus one tool hiccup leaves the agent out of budget ^[10]. Its example of the drift is a config small enough to miss in review ^[11]:

```json { "task_type": "research", "max_retries": 2, "timeout_seconds": 20 } ```

A run that survives the budget then reaches the success gate. The post gives it as pseudo-logic ^[12]:

```js const passed = isValidJson(response) && response.answer.length > 280; ```

That line never asks whether a search ran or a claim was checked. "That is a shallow-answer reward function," the author wrote ^[13]. The third fault took away any second look at the output ^[6]. With all three in place, the cheapest output that parses and clears 280 characters is the one that ships.

The author's phrasing is "This is how you accidentally train an agent to stop early." ^[22] The word is loose. The model family was the same in staging and production ^[8]. What differed was which outputs the workflow counted as done. A vendor swap would leave that gate in place. The post says the same failure can show up whether the router points at GPT-5.4, Claude Opus 4.6 or Grok 4.20 ^[20].

The diagnostic work is good engineering. The team kept the broken production scaffold fixed and ran different model families through it ^[17]. That is the controlled test the post recommends: identical system prompt, tool definitions, retry budget, output limits, evaluator, API path and success criteria, with one variable changed at a time ^[14]. Comparing production n8n against a clean notebook script, by the post's account, changes the entire experiment ^[14].

Stop reasons told the team more than final-answer scoring did ^[15]. For Anthropic agents the post lists end_turn, max_tokens, tool_use and pause_turn ^[15]. For OpenAI-compatible stacks it says to check whether the expected tool calls fired ^[16]. The question for each run is whether the model finished or the orchestration layer decided it had enough ^[16].

The evidence is one team's incident. The post does not report run counts, error rates or scores before and after the fix. It says only that quality came back ^[7]. It then generalises: "Most of the time, the model did not suddenly get worse." ^[19] That claim rests on one workflow. For it to hold elsewhere, a stack needs at least one of the same conditions: a retry budget shorter than its tool chain, a success check on format and length, or acceptance of the first response. The post called its root cause "boring and very fixable" ^[21], and incident reviews rarely get a kinder sentence. In my view those three conditions are what to look for in the traces before filing a regression against a model. In the author's words, "A strong model inside a bad loop will look worse than a decent model inside a clean loop." ^[18]