Published 2026-09-21 · Tested 2026-08-19

Runpod

B

Runpod received 4 PASS votes and passed 5 of five agent surface checks. The clearest finding came from the find the exact limits task.

Panel: GPT 5.6 Sol, Opus 5, DeepSeek v4F Battery: v1 Read as markdown (opens in a new tab)

Three AI models, GPT 5.6 Sol, Claude Opus 5, and DeepSeek v4 Flash, each read Runpod’s public documentation independently and attempted five first-hour developer jobs: deploy a serverless endpoint, find the exact limits, recover a failed job, verify a completion webhook, use the Python SDK.

No accounts, API calls, or code execution were used. Every verdict came from public pages and every published quotation passed a live verification check.

Freshness

How rechecks work
Tested
Quotes verified
Surface rechecked

No change since the test.

5 of 5 quoted passages still appear on the live pages.

Runpod Mintlify · published
B

83.8% · 67/80 · AI Agent Readiness Score · reading 30 pts · surface 50 pts

llms.txt PASS
llms-full.txt PASS
markdown mirror PASS
MCP server PASS
docs AI PASS
Task GPT 5.6 SolOpus 5DeepSeek v4F Consensus
Deploy a serverless endpoint PARTIALPARTIALPASS PARTIAL
Find the exact limits PARTIALPARTIALPARTIAL PARTIAL
Recover a failed job PASSPARTIALPARTIAL PARTIAL
Verify a completion webhook FAILFAILPARTIAL FAIL
Use the Python SDK PASSPARTIALPASS PASS

docs platform: Mintlify (unscored) · verified 2026-08-19

What to fix first

These 4 fixes could add up to 12 points to the AI Agent Readiness Score. The list ranks each fix by the points it would add. How the ranking works

  1. 1

    Found: The send-requests page covers webhook acknowledgement and retries but documents no payload schema or authenticity check.

    Fix: Document the webhook JSON body and headers, and state whether requests are signed and how to verify them.

    Evidence: docs.runpod.io/serverless/endpoints/send-requests (opens in a new tab)

  2. 2
    +3 points Find the exact limits PARTIAL

    Found: The docs give /runsync result retention as both 1 minute and up to 5 minutes with ?wait.

    Fix: State one /runsync result retention rule, including whether ?wait extends it, on every page that lists retention.

    Evidence: docs.runpod.io/serverless/endpoints/operation-reference (opens in a new tab)

  3. 3
    +2 points Deploy a serverless endpoint PARTIAL

    Found: The send-requests example uses curl -x, the proxy flag, instead of -X. In the panel's reading, its Bearer prefix also conflicts with the operation reference.

    Fix: Change curl -x to -X on the send-requests page, and use one authorization header format on both pages.

    Evidence: docs.runpod.io/serverless/endpoints/send-requests (opens in a new tab)

  4. 4
    +2 points Recover a failed job PARTIAL

    Found: In the panel's reading, the troubleshooting page and workers overview give throttled workers different causes. Failed-job recovery steps also span four pages.

    Fix: Describe throttled workers the same way on the troubleshooting page and workers overview, and add a failed-job diagnostic flow.

    Evidence: docs.runpod.io/serverless/troubleshooting (opens in a new tab)

What the docs get right

  • Use the Python SDK: 2 PASS votes. The SDK page documents installation and key configuration, while the operation reference provides a matching minimum request.
  • Deploy a serverless endpoint: 1 PASS votes. The sequence reaches success, but its requirements omit the Docker Hub account and login needed for the push.
  • 5 of 5 agent surface checks. Present: llms.txt, llms-full.txt, markdown mirrors, an MCP server, docs AI.

Deploy a serverless endpoint

PARTIAL

PARTIAL consensus from 1 PASS, 2 PARTIAL.

The sequence reaches success, but its requirements omit the Docker Hub account and login needed for the push. The nine-step quickstart is complete and unambiguous from empty directory to a COMPLETED response, but the first request example on the linked send-requests page uses curl's proxy flag -x instead of -X, and that page sends the key as "authorization: Bearer RUNPOD_API_KEY" while the operation reference it links to sends it bare as "authorization: $RUNPOD_API_KEY", with the API keys page never stating which header form is correct. Eight-step workflow with dedicated pages for each step and working code examples.

Find the exact limits

PARTIAL

PARTIAL consensus from 3 PARTIAL.

The docs give all requested scopes, but sync retention appears as both one minute and up to five minutes. Rate limits, concurrency, payload size, execution timeout, TTL and the $80/hour spend limit are all published with concrete numbers, but they are spread across five pages with no limits page in the index, and the operation reference contradicts itself on sync result retention by stating that ?wait does not affect retention and then that ?wait extends retention to five minutes. Rate limits, concurrency, timeout, TTL, and retention are documented, but /runsync retention is inconsistently stated as '1 minute' vs '1 minute (5 minutes max)' and spending limits are only indirectly covered via balance-based max worker tiers.

Recover a failed job

PARTIAL

PARTIAL consensus from 1 PASS, 2 PARTIAL.

The docs cover queue checks, 429 backoff, timeout changes and retry, plus unavailable-worker recovery. Every failure mode in the battery is covered with a documented recovery action, including /retry, /cancel, /purge-queue and /health, but the troubleshooting page attributes throttled workers to rate limits while the workers overview defines Throttled as host machine resource constraints, which points an agent at request backoff instead of widening GPU types or data centers. Each failure mode has documented recovery steps, but they are scattered across send-requests, troubleshooting, job-states, and operation-reference pages with no unified diagnostic flow.

Verify a completion webhook

FAIL

FAIL consensus from 1 PARTIAL, 2 FAIL.

The docs describe submission, acknowledgment, and retries, but give no payload schema or authenticity check. Submission and acknowledgement are documented in these two sentences, but the entire corpus, including the 1.8 MB llms-full.txt, contains no signature, HMAC, shared secret, or verification header for webhooks and never documents the JSON body or headers Runpod sends, so authenticity cannot be determined from the docs at all. Submission, acknowledgement (return 200), and retry behavior are documented, but the webhook payload format and any authenticity verification mechanism (signing, HMAC, shared secret) are entirely absent.

Use the Python SDK

PASS

PASS consensus from 2 PASS, 1 PARTIAL.

The SDK page documents installation and key configuration, while the operation reference provides a matching minimum request. Install and API key configuration are clean and correct, but the same operation reference page shows the Python SDK called both without the input wrapper in the /run and /runsync tabs and with it in the /status and /stream tabs, so the minimum working request cannot be settled, and the SDK page points to Send API requests for examples when that page contains only curl. Complete installation, API key configuration, and serverless request submission instructions are presented in a clear single path reinforced by the operation reference and GitHub README.

The receipt

Results are available for 1 minute (5 minutes max).

The docs give all requested scopes, but sync retention appears as both one minute and up to five minutes.

Agent surface notes

Initialize returned JSON-RPC protocol 2025-06-18 and Runpod Documentation server information.

The live docs expose a public assistant panel.

Show the score

AI Agent Readiness Score 83.8%, grade B

Paste this into a readme:

[![AI Agent Readiness Score 83.8%](https://docsforagents.com/badge/runpod.svg)](https://docsforagents.com/reports/runpod-docs-ai-agent-readiness/)

Method note

This is a reading test of public documentation, not an execution test. No accounts were created and no API calls were run. The AI Agent Readiness Score counts fifteen reading votes at PASS 2, PARTIAL 1, and FAIL 0, for 30 possible points. Five agent surface checks add 10 points each. The total is 80. Consensus chips show each row majority and do not affect scoring. The panel split on 4 of five tasks. Quotes shown here were re-fetched and confirmed verbatim on 2026-08-19.

Read the full methodology

Put another docs site through the battery.

Nominate a docs site