Agentic AI · 2026
Relayd
relayd.io · product I built · Founder and engineer
An orchestrator for long-running agent workflows: durable state, typed tool contracts, compensation on failure, and human approval steps that survive a redeploy mid-run.
agent runs per month
control-plane availability
runs lost to deploys since launch
The problem
Every agent framework I tried assumed the run finishes in one process. Real workflows wait on a human for two days, call a flaky vendor API, and need to resume exactly where they stopped — that is a durable-execution problem, not a prompting problem.
What I did
- Modelled each run as a state machine persisted per step, so a deploy or a crash resumes rather than restarts.
- Gave every tool a typed contract with timeouts, retry policy and a compensating action, validated at registration.
- Built an approval primitive: the run parks, notifies, and continues from the human decision with full audit trail.
- Added a replay console — any historical run can be re-executed against a new prompt or model to see what would change.
Architecture
Spring Boot control plane on EKS, workflow state in Postgres with an outbox pattern onto Kafka, workers scaled by queue depth via KEDA, and an OpenTelemetry trace per run stitched through every tool call. Self-hostable via Helm chart.
Request path
Similar problem on your side?