In static pipeline frameworks like Airflow or Argo Workflows, the execution graph is fixed at compile time. In Asya, routes are data embedded in each message, and actors can rewrite the future at runtime.

Routes are data, not infrastructure#

Every envelope carries a route object with three fields:

• prev — actors already visited (read-only)
• curr — the actor currently processing this message (read-only)
• next — the actors still to visit (writable)

Because the route is inside the message, different messages can follow different paths through the same set of actors.

Actors can change the future#

A generator handler can insert actors at the front of route.next using the slice notation .route.next[:0]. This prepends without losing the remaining route:

def handler(payload: dict):
    confidence = payload["score"]
    if confidence < 0.7:
        # Insert human-review before whatever comes next
        yield "SET", ".route.next[:0]", ["human-review"]
    else:
        yield "SET", ".route.next[:0]", ["auto-approve"]
    yield payload

Avoid overwriting .route.next entirely (e.g., yield "SET", ".route.next", [...]). This discards the remaining route and breaks downstream actors that expect to receive the envelope. Use .route.next[:0] to prepend safely.

Only .route.next is writable. .route.prev and .route.curr are read-only, enforcing a forward-only execution model — actors cannot rewrite history.

What this enables#

• LLM judges — a model evaluates quality and routes to different downstream actors based on the result
• Confidence-based branching — low-confidence predictions go to human review
• Agentic loops — an actor routes back to itself (via next) for iterative refinement
• Human-in-the-loop — route to x-pause to checkpoint and wait for human input, then resume with x-resume

Contrast with static DAGs#

Further reading#

• ABI Protocol — full yield protocol for GET/SET/FLY operations
• Agentic Patterns — routing patterns for AI agent workflows
• Pause/Resume — human-in-the-loop via x-pause and x-resume