Execution Chain Taxonomy.
# Abstract
This RFC defines the Execution Chain Taxonomy (ECT), a technical classification system for agentic workflows. By standardizing the levels of logic persistence and error isolation, ECT provides a framework for building deterministic AI infrastructure at scale.
# Execution Levels
Level 1: Stateless (Atomic). Level 2: Stateful (Iterative). Level 3: Validated (Multi-Agent). Level 4: Hardened (Infrastructure-Integrated).
# Orchestration Standard
ECT-compliant systems MUST implement clear transition gates between execution nodes and SHOULD utilize a centralized state manager for complex logical branching.
# Architectural Blueprint
{
"rfc": "002",
"pattern": "VALIDATED_MULTI_AGENT",
"constraints": {
"node_isolation": true,
"state_persistence": "REDIS_STORE",
"timeout_ms": 30000,
"max_retries": 2
},
"topology": "HUB_AND_SPOKE"
}Cite ECT Standard
@techreport{sable2026rfc002,
author = {Sable, Datta},
title = {RFC-002: Execution Chain Taxonomy Standard v2.1.0},
institution = {Surgical AI Workspace},
year = {2026},
url = {https://dattasable.com/knowledge/rfc/002-execution-chains}
}Internal reference · Illustrative example
Example: resume a report without sending it twice
A weekly report chain can load data, compute figures, draft commentary, validate it, and request publication approval. Store the run identifier and status after each completed stage. If the draft stage times out, retry that stage using recorded inputs rather than reloading different data or publishing the report again.
Verification checklist
- Use a stable reporting period and source snapshot.
- Make externally visible actions safe against duplicate retries.
- Reject commentary that conflicts with computed figures.
- Stop when a required source is unavailable.
This RFC proposes local conventions for state and transitions. A fixed path controls execution order; it does not make generated commentary factually correct. Test interruptions at stage boundaries, and keep approval separate from generation when publication has external consequences.
Evaluate workflow recovery and agent stopping rules