AI harness for reliable, long-running coding-agent tasks
4.8(6)6Artificial Intelligence
Relay is an AI harness for coding agents built for long-running tasks. It supervises agents while you are away, resumes interrupted sessions, and recovers from failures. Relay runs on your machine with your existing repository and toolchain. It reads repository conventions, remembers session settings, runs configured tests, and produces signed receipts you can reopen and verify. Relay provides evidence about the configured checks; developers still review the code and decide what ships.
• Keeps long-running coding-agent sessions moving
• Resumes after interruptions and recoverable failures
• Uses your existing repository conventions and local toolchain
• Runs configured tests and verification checks
• Produces signed, re-checkable receipts
• Supports solo and team workflows
• Leave substantial coding tasks running unattended
• Recover interrupted agent sessions without starting over
• Validate changes with repository-native tests
• Review evidence before merging or shipping
• Maintain consistent workflows across developers and agents
Comments
6small indie software guys
Having both sides, job seekers and recruiters, in one platform is interesting. Does the ATS-friendly resume feature show the candidate why a resume would get filtered out, or does it just rewrite it? Seeing the reason would make it a learning tool, not just a fix.
small indie software guys
The signed receipts you can reopen later are the part that stands out to me. Being able to come back and prove which checks ran while I was away is the trust problem with long agent runs. How does Relay handle a task that's halfway through a migration when it gets interrupted? Does it roll back or pick up from the last checkpoint?
small indie software guys
sorry wrong product, disregard!
Signed receipts are a strong idea for team review. Can a teammate verify a receipt on their own machine without access to the original agent session, and how does Relay record a configured test that is flaky rather than clearly failing?
Founder of AgntFlow
Love that it re-runs tests and lint before opening the PR. That’s what makes unattended agents actually usable. Does the learning between runs stick per repo?
Signed receipts for unattended runs make a lot of sense. One failure I've hit running several coding agents in parallel: two sessions on separate branches each picked the same database migration number, and each run passed its own tests. The clash only showed up after merging. Does Relay's verification run against the merged state, or only inside each session? A check like that is the one I'd most want a receipt for.