Logo SVG copied to clipboard

Your failing requests now fix themselves

Diagram of the Manifest gateway: a request fails with a 400 temperature out of range error, Auto-fix applies a patch and the retried request returns 200 OK

Auto-fix is live. If you joined the waitlist, it’s already running on your account. You have nothing to turn on.

Errors were eating your agents

Here’s a night we’ve all had. Your agent runs a long workflow, and at some point the provider renames a parameter: max_output_tokens becomes max_completion_tokens on the model you pinned months ago. The provider answers 400 and your agent dies mid-run. You find out in the morning, and the fix is a one-line rename the provider decided for you.

Deprecated models and unsupported fields kill agents the same way. Every one of these errors burns hours while nothing runs.

We got tired of watching that happen. Manifest now repairs these requests itself.

A 400 error response from a provider: unsupported value, temperature does not support 0.2 with this model

The dashboard now counts what matters

The Overview got rebuilt around one question: how are your requests doing? It shows how many ran, how many failed and how many were recovered. The same reading exists at both levels: global for your whole workspace, and one per agent.

The Requests log follows. You can filter by connection, recovery method or period. So you can easily see, for example, everything Auto-fix saved this month.

The Manifest Overview dashboard showing success rate, recovered requests, requests recovered by Auto-fix and by Fallback, failed requests and the recovered requests chart

Messages become requests

Our gateway now logs requests. A request can hold one to several attempts, the actual calls sent to providers. When an attempt fails, the next one steps in.

Each row shows one request. Click it to see every attempt related to it in detail: the error the provider returned and the exact change Auto-fix applied. So you can also repair the cause on your side, and that error never comes back.

How Auto-fix works

When a provider rejects a request with a fixable error, the router applies a patch and sends the request again. Each patch targets a known error and applies a known correction.

Today Auto-fix covers a first set of errors. That set grows with usage, as new errors come in.

An auto-fix detail showing a failed attempt retried as attempt 2 with the applied patch: rename_param max_tokens to max_output_tokens

Where Auto-fix runs

Auto-fix is available on Manifest Cloud, free plan included. The patches come from errors seen across cloud traffic.

Tell us everything

We opened this beta to learn from you. In a few days we’ll reach out for your first impressions. Don’t wait for us if you want to chat: just contact us.

Try it now

If you received the announcement email, your account already has Auto-fix. Open your dashboard and look at your Requests log: some repairs may already be there. If you didn’t, you can ask us for a demo here.

Start routing for free. Scale with your team.