Mastra•5d ago

1 step Agent Taking up to 15 minutes

Hi team, I am trying to debug a performance issue. I have a single turn agent sometimes taking up to 15 minutes to respond to a single structured ouput prompt well under 100k total tokens in size.

I am trying to figure out how to speed it up. Both gpt-5-mini with minimal reasoning effort and gpt-4.1-mini have had timeout issues or insane response times. What is the best way to diagnose the issue?

Chad HolmesOP•12/12/25, 4:18 AM

[
{
"min_seconds": 4.2,
"p25_seconds": 65.24,
"p50_median_seconds": 143.72,
"p75_seconds": 249.12,
"p90_seconds": 363.14,
"p95_seconds": 588.7,
"p99_seconds": 890.03,
"max_seconds": 944.81,
"avg_seconds": 191.61
}
]

Chad HolmesOP•12/12/25, 4:35 AM

Nevermind, looks to be an openAI issue. The models really do struggle with any complexity in the structured output huh :/

dns•12/12/25, 4:47 AM

I've also seen delays with structured outputs

Chad HolmesOP•12/12/25, 5:37 AM

Just ran a test with gpt 4.1 mini

FLAT Schema (simple prompt):
Avg: 2.88s
Range: 2.55s - 3.27s

COMPLEX Schema (simple prompt):
Avg: 7.30s
Range: 6.38s - 7.99s

PRODUCTION Schema (large prompt):
Avg: 56.69s
Range: 35.95s - 77.44s

FLATTENED PRODUCTION Schema (large prompt):
Avg: 31.21s
Range: 30.56s - 31.86s

dns•12/12/25, 6:16 AM

Using Claude, I've had really bad experiences with structured outputs. I gave 3 tools to an agent and told it to getOrders, and it tried to createEvent.. and it has a very hard time even building the structured output response

And using Open AI, I always get zod validation errors coming from vercel-ai SDK... The only way for me to bypass was to use OpenAI's response format directly (openai: { response_format: { type: 'json_object' } })... rather than relying on mastra's structuredOutput

dns•12/12/25, 6:17 AM

Are you using response_format: { type: 'json_object' } or mastra's structuredOutput?

Chad HolmesOP•12/12/25, 6:26 AM

Im using mastra's structuredOutput

dns•12/12/25, 6:26 AM

Which mastra version do you have?

Chad HolmesOP•12/12/25, 6:38 AM

"^0.18.6"

dns•12/12/25, 6:48 AM

Yep. Try upgrading and it'll probably break. At least that's what happened to me

Sebastian•12/12/25, 7:54 AM

I don't know if this makes a difference, but I've seen the structured output schema repeated in the prompt in several prompts inside Mastra itself (see for example here).
No idea what the motivation was here (might be a compatibility thing from the past). Maybe worth a try to mimic that.

GitHub

mastra/packages/evals/src/scorers/llm/prompt-alignment/prompts.ts a...

The TypeScript AI agent framework.

Assistants, RAG, observability. Supports any LLM: GPT-4, Claude, Gemini, Llama. - mastra-ai/mastra

Mastra Triager•12/12/25, 12:03 PM

Created GitHub issue: https://github.com/mastra-ai/mastra/issues/11126

If you're experiencing an error, please provide a minimal reproducible example whenever possible to help us resolve it quickly.

Thank you for helping us improve Mastra!

SSebastian I don't know if this makes a difference, but I've seen the structured output sch...

caleb•12/12/25, 4:49 PM

The schema for those evals/scorers was probably included in the prompt because we don't know what model you will use or if it even supports structured outputs. The eval could be run with any model so this probably increases reliablility for some models.

You can also use structuredOutput.jsonPromptInjection = truestructuredOutput.jsonPromptInjection = true and it will add your schema into the system prompt instead of using response_format: {type: "json_schema", ...}response_format: {type: "json_schema", ...}

structuredOutput: {
    schema: z.object({
      hello: z.string(),
    }),
    jsonPromptInjection: true,
  }

structuredOutput: {
    schema: z.object({
      hello: z.string(),
    }),
    jsonPromptInjection: true,
  }

https://mastra.ai/docs/agents/overview#response-format

Using Agents | Agents | Mastra Docs

Overview of agents in Mastra, detailing their capabilities and how they interact with tools, workflows, and external systems.