01
Learn
Your model fixes a failing test. Relay keeps the fix, then proves it in a sandbox: it has to transfer to renamed and restructured variants with zero regressions before it goes live.
Gated before it ships
Fix it once.Never twice.
Verified fixes forcoding agents, minus the model call
Relay learns the fixes your agent's model makes and replays them on the next similar failure. Private beta.
Every coding agent pays its model to rediscover fixes it has already made. Relay remembers them, checks them, and hands them back for free. Here's how:
Not a cache. A memory for fixes.
A cache replays the same request. Relay carries what your model learned to failures it has never seen before.
02
Transfer
The next time a similar failure shows up, in another file, with other names, Relay writes the fix itself. Your agent never calls the model for that turn.
43 of 50 replay tasks, zero tokens
03
Verify
Your tests are the judge. Misses fall through to your model as usual, free. A patch that fails is never billed, and every promotion can be rolled back.
0 bad fixes in the replay
Your agent keeps paying for the same fix.
The problem
Coding agents hit the same failures over and over: the off-by-one in a pager, the empty queue that returns undefined, the null nobody checked. Each time, the agent sends the whole context back to the model and pays for it to think the fix through from scratch.
The model gets it right. It got it right last week too. Nothing about that first fix is kept, so the bill grows with every task and the agent never gets faster at problems it has already solved.
Relay keeps the fix, proves it, and makes the next one free.
Models are brilliant. They just don't remember.
What Relay changes
A frontier model is the right tool for a failure nobody has seen before. It is an expensive way to fix the fiftieth copy of one you already have.
Relay sits in front of the model. When a turn matches a fix it has learned and verified, Relay writes the patch itself in milliseconds. Everything else goes to your model exactly as it does today.
Your model, for
Failures it has never seen, and anything Relay isn't sure about.
No longer paying for
The same fix again, at full price, every time it comes back.
If Relay is ever down or unsure, it steps aside. Your agent always runs.
Watch. Learn.
Transfer. Verify.
In a real replay from an empty library, the model wrote 7 of 50 fixes. Relay wrote the other 43, and every task still passed.
01Watch
Relay sees your agent's normal traffic: the failing code, and the fix that turned the test green.
02Learn
The fix becomes a capability only after it passes every gate in a sandbox.
03Transfer
A similar failure arrives. Relay writes the patch itself, with no model call.
04Verify
Your tests decide. You pay only for fixes that pass.
Same 50 tasks. 70% fewer model calls.
Real-agent replay
Measured, not projected. We replayed a real coding agent through Relay from an empty library: 250 model calls became 71, and 6.4 million tokens became 2.0 million. All 50 tasks were solved, and not one Relay fix failed and fell back to the model.
We ran it on three models, GPT, Gemini and an open-weight model. All three landed between 69% and 72% saved. Underneath that number are four promises:
How it stays safe01Fail-open
Any Relay error is a miss. Your agent's normal path always runs.
02Shadow mode
Start by measuring what would have hit on live traffic. Nothing is served.
03Yours only
Fixes learned in your project are only ever served to your project.
04Reversible
Every promotion is logged. Retire one, or stop everything, in a click.
One line to plug in. Nothing to break.
Integration
Relay speaks the OpenAI API. Change your client's base_url to Relay, keep your own provider key, and every call goes through exactly as before. Relay only steps in on the turns it can prove it knows.
Rolling it out is three steps, and you can stop at any of them:
- 01Point. Set base_url to Relay. Chat Completions or Responses, streaming included.
- 02Shadow. Relay records what it would have answered on live traffic, and serves nothing.
- 03Active. Relay serves verified fixes. Misses still go to your model, free.
You pay per verified fix. Misses and shadow decisions cost nothing.
Works with the agent you already run.
Relay is a drop-in gateway, not a new framework. If your agent talks to an OpenAI-compatible endpoint, it already works with Relay.
Codex CLI
Point Codex at Relay and run it as usual. Relay answers the turns it already knows over the Responses API.
OpenAI SDK
Any client that takes a base_url. Same calls, same models, your own provider key.
Claude
Route Claude through the same Chat Completions endpoint. Relay forwards with your Anthropic key.
Gemini and open models
Any OpenAI-compatible provider. The replay landed between 69% and 72% saved on all three models we tried.