Dual-Lane Architecture
OpRelay separates work into two lanes with fundamentally different execution profiles.
Human Lane
The human lane is for work that benefits from your judgment — architecture decisions, escalations, code review, and anything where context and deliberation matter more than speed.
| Property | Value |
|---|---|
| Context window | Full |
| Model tier | Opus-class |
| Approval | Required before completion |
| Throughput | ~0.3 tasks/min |
| Use cases | Architecture, review, escalation |
Agent Lane
The agent lane is for autonomous execution — implementation, testing, iteration. Agents work at machine speed with compressed context and lighter-weight models.
| Property | Value |
|---|---|
| Context window | Compressed |
| Model tier | Haiku / Sonnet |
| Approval | Not required |
| Throughput | ~2.4 tasks/min |
| Use cases | Implementation, testing, CI |
How they interact
The broker routes work to the appropriate lane based on configuration. Agents in the agent lane can escalate to the human lane when they hit decisions beyond their scope. Humans can release tasks back to the agent lane after providing direction.
This creates a natural flow: humans set intent, agents execute, and the system coordinates the handoff.
Pipeline model
Within the agent lane, work flows through a three-stage pipeline:
Supervisor → Worker → Reviewer
- Supervisor — decomposes work, sets priorities, manages the backlog
- Worker — executes tasks in isolated containers
- Reviewer — validates output, approves or sends back
The pipeline is prescribed, not emergent. Each stage has defined responsibilities and the broker enforces the flow.