Article

Prepare ElevenAgents apps for November 26 API timeouts

Teams using ElevenAgents can check existing retries, separate client and server timeouts, and handle uncertain call outcomes before November 26.

Editorial illustration for Prepare ElevenAgents apps for November 26 API timeouts: a controlled task flows from input to output. Not documentary evidence.

ElevenLabs will enforce server-side request timeouts for its ElevenAgents API from November 26, 2026. Its November 16 notice says requests may return HTTP 504 during temporary slowdowns. Teams running voice-agent integrations should check existing retry handling—and verify uncertain outcomes before resending requests that start calls.

An integration that already handles these cases may need no code changes. The task is to check both whether the client tries again and whether repeating that particular action is safe.

Find the affected requests

The notice names /v1/convai. Inventory the clients using that API, including background jobs and workflow connectors. Record the installed library version, request method and any retry settings. For a managed connector, ask its maintainer how it handles a 504 and prevents an uncertain action from being blindly repeated.

If the same backend also uses Claude, keep that provider’s SDK/version review separate. RohitAI’s Claude API compatibility checklist covers that different integration change.

Check the retries you already have

The official Node SDK documentation describes two automatic retries with exponential backoff—waiting progressively longer between attempts. The pinned retry implementation includes 504 responses: one initial attempt plus up to two retries. Check your installed version and overrides; this is not a statement about Python, browser SDKs or third-party connectors.

Review retries in the SDK, gateway and job runner together. As a conditional calculation, if an outer workflow retries the entire SDK invocation twice, and each invocation exhausts two SDK retries, the ceiling is (1 + 2) × (1 + 2) = 9 HTTP attempts. That is not nine guaranteed phone calls or a billing estimate. Set a coordinated limit on attempts and total waiting time before adding another loop.

The retry helper checks response status; it does not decide whether repeating the business action is safe. A working retry mechanism is only half the preparation.

Check before repeating a call

Under HTTP’s idempotency rules, a request is idempotent when repeating it has the same intended server effect as sending it once. Read-only GET requests fit that model. Automatically repeat a POST that creates a resource only when its design makes repetition safe or you can establish that the first attempt was not applied.

A 504 means a gateway did not get an upstream response in time. It does not establish whether the upstream action completed. The ElevenLabs notice does not say that timed-out operations are cancelled. Treat a timed-out call-starting request as unresolved until you can verify what happened.

The outbound-call reference does not document an idempotency-key contract or a guarantee against duplicate execution. Do not invent such a guarantee by adding an arbitrary header. This is a limit of the published reference, not proof that the service has no internal safeguards.

For an action whose repeat behavior is unverified, the pinned Node Twilio client accepts maxRetries: 0 in its request options to disable automatic retries at that SDK layer. Pair that choice with an outcome-checking process, and inspect outer retries too. Disabling retries does not resolve an uncertain result.

Preserve available conversation or call identifiers. The API supports conversation lookup by ID and listing conversations by agent and start time, which can help investigate. An empty immediate search is not documented proof that nothing happened. Keep unresolved work pending for investigation instead of automatically placing another call.

Keep client and server timeouts separate

A client timeout controls how long your app waits; it cannot extend ElevenLabs’ server deadline. The notice supplies neither a numerical server limit nor an enforcement hour or time zone. There is no supported universal timeout value to copy into every integration.

Even client defaults need checking: the README describes 60 seconds, while the pinned Twilio outbound-call method falls back to 240 seconds unless request or client options override it. Neither number describes the new server cutoff. Inspect the method in your installed package before changing timeoutInSeconds.

The Node fetcher handles HTTP error responses and client timeout/abort failures separately. Do not assume a 504 retry rule covers both. Keep rate limiting distinct too; teams also using SiliconFlow can consult its separate China-platform rate-limit and 429 guide.

Verify recovery before November 26

Use isolated mock responses to check the intended behavior before rollout:

  • A safe read receives one 504 and then succeeds: confirm the retry count and waiting time.

  • Repeated 504s or a client abort: confirm the app stops within its total time budget and displays an accurate status.

  • A call-starting request has an unknown outcome: confirm neither the SDK nor the job runner blindly resends it, and that someone can investigate.

These are suggested checks, not reported test results. Do not place real customer calls merely to exercise retry logic. After enforcement, monitor 504s, total attempts and unresolved actions—not only whether the last HTTP request succeeded.

The timeout notice announces no price change or required plan upgrade. Agents pricing charges for usage, with model usage additional and external telephony billed separately. A 504 alone cannot tell you what usage occurred; check actual usage records rather than treating every HTTP attempt as a billable conversation.