Agentic Graph Dispatch

Directed acyclic graphs orchestrating multi-agent workflows with dependency resolution, parallel execution and checkpoint-based resumption.

Scheduling

The graph decides what can run at once

Units are ordered topologically, then anything independent is released together. Parallelism falls out of the dependency structure rather than being guessed at.

wave plan
wave w1 ├─ u1 contracts ready
├─ u2 migration ready
└─ u3 api surface ready
wave w2 └─ u4 consumers blocked on u1, u3
✔ 3 of 4 units run concurrently

Isolation

Parallel builders never collide

Before a wave opens, the planner checks file overlap across its units. Two agents are never handed the same file, so parallelism costs nothing in merge conflicts.

overlap pre-check
units in wave4
files touched23
overlapping paths0
worktrees4 isolated
verdictsafe to parallelise

Recovery

A failed node resumes, it does not restart

Each node checkpoints on completion. When one fails, the graph replays from that node forward and everything already proven stays proven.

resume
▶ replaying wave w2 from checkpoint
u1 contracts cached
u2 migration cached
u4 consumers re-running
✔ 2 of 3 units skipped

DAG execution

Dependency-aware scheduling with topological ordering.

Parallel branches

Independent subgraphs run concurrently for throughput.

Checkpoint resume

Retry from the last good node, never from the top.