Concepts/Concept
ADPS Engineering Concept Registry
Task DAG and State Machine
Make dependencies, node status, acceptance, and recovery explicit.
Application context
Payroll setup must create a snapshot before importing a template, associate employees only after the payroll group exists, and roll back when import fails. Describing this order in a Skill leaves the model responsible for remembering every dependency during execution.
Definition
The task DAG stores dependency edges. The node state machine stores lifecycle transitions such as pending, ready, running, waiting, completed, failed, and compensating. A node becomes schedulable only when its dependencies and preconditions hold.
Engineering mechanism
The Planner creates typed nodes with inputs, outputs, resource keys, authority, acceptance, retry, and compensation. The Executor claims a ready node through a lease. The Verifier records acceptance evidence before completion. Every transition is conditional and persisted; recovery reconstructs runnable nodes from the graph and external receipts rather than replaying conversation.
Boundary
A short, reversible sequence can remain a prompt chain or ReAct loop. Strict dependencies, parallel branches, approval waits, rollback, and cross-session recovery justify a DAG and state machine. The graph controls execution order; it does not decide whether the original business goal was correct.
Proposal and provenance
- Concept group
- Planning, execution, and write-back
- Initial source within ADPS
- Initial practice: Bo Liang
- First recorded in
- Dongfang Yiteng execution-agent case ()
- ADPS editorial work
- ADPS placed dependency, acceptance, and recovery in one agent task contract.
- Current standing
- Agent adaptation of established terms
Initial source: Dongfang Yiteng Execution Agent case report; contributed by Bo Liang.
Chronicle
- Recorded source
- Dongfang Yiteng execution-agent case ()
- Source date
- First published on ADPS