Registry-Governed Agent Lifecycle Completing EDDOps with Evaluation-Driven Registration, Promotion, and Retirement on AWS

Jul 16, 2026

Upload by Finn

Evaluation-Driven Development and Operations (EDDOps) treats evaluation as a continuous governing function for LLM agents, but prior approaches use the agent registry as a passive catalog. This work argues that the registry must act as the active control plane across the full lifecycle: registration, evaluation-driven promotion, MCP-native discovery, version management, and retirement. Four contributions formalize an evaluation-gated lifecycle state machine, implement registry-as-control-plane on AWS Bedrock AgentCore with multiple agents, validate complete lifecycle behavior, and introduce a cost-to-performance framework including lifecycle costs.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

Introduction

Motivating Problem

Contributions

Scope and Methodology

Registry-Governed Agent Lifecycle: Completing EDDOps with Evaluation-Driven Registration, Promotion, and Retirement on AWS AgentCore Dr. Richard Kang DoiT International [richard@doit.com](richard@doit.com) Vincent Wang arXiv :2607 .00345v 1 [ cs . SE] 1 Jul 2026 Abstract—Evaluation-Driven Development and Operations (EDDOps) establishes evaluation as a continuous governing function for LLM agents, yet existing implementations treat the agent registry as a passive catalog—a terminal artifact that records deployment decisions already made. This paper argues that the registry must instead serve as the active control plane governing the full agent lifecycle: registration, evaluation-driven promotion, MCP-native discovery, version management, and retirement. We present four contributions: (1) a lifecycle state machine formalization where registry status transitions are gated exclusively by evaluation evidence; (2) an architectural instantiation on AWS Bedrock AgentCore demonstrating registry-as-control-plane with six agents (three managed, three BYO) progressing through DRAFT → APPROVED → PUBLISHED → DEPRECATED → RETIRED states; (3) empirical validation through a proof-ofconcept exercising the complete lifecycle—including stale-agent detection, automated re-evaluation, score-driven promotion/demotion, and MCP-based agent-to-agent discovery; and (4) a costto-performance framework incorporating both per-interaction cost and lifecycle cost (evaluation overhead, registry operations, version proliferation). The instantiation demonstrates that when the registry governs lifecycle transitions rather than merely recording them, EDDOps achieves its full promise: agents are born evaluated, live under continuous evaluation, and retire when evaluation evidence warrants—closing the gap between evaluation research and operational governance. Index Terms—EDDOps, Agent Registry, Agent Lifecycle, LLM Agents, Evaluation-Driven Development, AWS AgentCore, MCP Discovery, Model Selection, AgentOps I. INTRODUCTION Enterprise adoption of LLM-based agents—systems that autonomously reason, plan, and execute multi-step workflows—requires evaluation methods that go beyond one-time benchmark scores [2], [4] . Organizations deploying agents for compliance-critical tasks need to answer not only which model delivers acceptable quality at justifiable cost, but also how agents should be governed across their full operational lifecycle: from initial registration through production deployment to eventual retirement. Xia et al. [1] introduced Evaluation-Driven Development and Operations (EDDOps), proposing a process model and reference architecture that embed evaluation as a first-class, lifecycle-spanning function. Their work identified six crosscutting evaluation drivers (D1–D6) and prescribed “governed registry-based discovery” as an architectural responsibility. However, the EDDOps reference architecture remains at the level of analytical generalization—it does not formalize how registry state transitions should be coupled to evaluation evidence, nor does it address the economic dimension of lifecycle governance. In practice, this means organizations deploy agents through ad-hoc approval workflows disconnected from evaluation infrastructure. Agent registries serve as passive catalogs—terminal artifacts that record deployment decisions already made—rather than active governors of agent lifecycle. This paper addresses both gaps by presenting a complete EDDOps instantiation on AWS Bedrock AgentCore [6] that elevates the Agent Registry from passive catalog to active control plane. We formalize a lifecycle state machine where every transition—from initial registration through production deployment to eventual retirement—is gated by evaluation evidence. The registry becomes the single source of truth not only for what agents exist, but for whether they should continue to exist based on continuous evaluation. A. Motivating Problem Consider an

FAQs

What problem does the paper identify with current EDDOps implementations?

It identifies that existing systems treat the agent registry as a passive catalog that records deployment decisions, rather than governing lifecycle transitions based on evaluation evidence.

How does the proposed lifecycle govern agent state transitions?

A lifecycle state machine is formalized so every transition (from registration through production deployment to retirement) is gated exclusively by evaluation evidence.

What is demonstrated in the AWS Bedrock AgentCore instantiation?

The registry is implemented as a control plane orchestrating evaluation scheduling, score-driven transitions, MCP-based discovery, and retirement across multiple agents progressing through lifecycle states.

Jul 16, 2026

Upload by Finn