The Agent Registry Becomes the Retirement Board
Richard Kang and Vincent Wang's arXiv paper Registry-Governed Agent Lifecycle: Completing EDDOps with Evaluation-Driven Registration, Promotion, and Retirement on AWS AgentCore argues that an enterprise agent registry should not be a catalog of decisions already made. It should be the place where agents are born, promoted, discovered, deprecated, and retired by evaluation evidence.
Registry as Governor
The paper is Registry-Governed Agent Lifecycle: Completing EDDOps with Evaluation-Driven Registration, Promotion, and Retirement on AWS AgentCore, arXiv:2607.00345 [cs.SE]. The arXiv API record lists Richard Kang and Vincent Wang as authors and records submission on July 1, 2026. The paper frames Evaluation-Driven Development and Operations, or EDDOps, as a continuous control loop for LLM agents rather than a one-time test before launch.
The local angle is simple: a registry is not neutral infrastructure. If it only records agents after approval, it becomes an obituary for governance decisions. Kang and Wang instead make the registry the control plane that decides whether an agent can advance, remain discoverable, be demoted, or retire.
That belongs beside agent runtime governance, AI agent observability, AI agent identity, and agent framework infrastructure. The missing governance object is not another benchmark score. It is the stateful record that binds scores, versions, owners, discovery, and retirement policy.
State Machine
The paper formalizes five states: DRAFT, APPROVED, PUBLISHED, DEPRECATED, and RETIRED. In this model, DRAFT means registered but not yet evaluation-cleared. APPROVED means initial gates have passed. PUBLISHED means production gates and human sign-off allow full discovery and service. DEPRECATED means score regression, staleness, or supersession has made the agent suspect. RETIRED means discovery and serving stop while the evaluation history remains preserved for audit.
The useful part is not the names. It is the transition rule. The paper makes evaluation evidence mandatory for state movement. A DRAFT agent must pass eligibility gates before APPROVED. APPROVED requires sustained gate compliance and human sign-off before PUBLISHED. PUBLISHED can move to DEPRECATED when scores regress, staleness exceeds policy, or a newer version passes gates. DEPRECATED can either recover through re-evaluation or retire after migration and grace-period conditions are met.
This turns model selection into lifecycle discipline. A cheaper model is not a bargain if it fails production gates. A once-strong agent is not forever approved if its tools, prompts, model, or operating context have drifted.
Discovery Is Policy
The registry paper treats discovery as a policy lever. If only PUBLISHED agents appear in Model Context Protocol discovery, then quality and safety state directly shape what other agents and developers can find. Discovery is no longer a search convenience. It becomes an enforcement surface.
AWS's Agent Registry preview post gives the product backdrop: the registry stores structured records for agents, tools, MCP servers, skills, and custom resources; supports registration through console, SDK, API, or endpoint-derived metadata; and applies an approval workflow before broad discoverability. The paper pushes that idea further by coupling discoverability to evaluation state, staleness policy, and retirement.
That coupling matters for agent sprawl. Enterprise agents are easy to copy, wrap, rename, and forget. A hardcoded endpoint can keep calling an obsolete agent long after its owner has moved on. A registry-mediated call can ask a different question first: is this agent currently eligible to be found?
Cost and Evidence
The paper validates the architecture as a proof of concept, not a broad benchmark. It uses a 15-case single-turn dataset, 9 multi-turn scenarios with mock tool data, six agents, three foundation models, and two deployment paths: managed runtime and bring-your-own deployment. The authors explicitly say production deployments need larger datasets, longer monitoring windows, and formal statistical testing.
Within that bounded exercise, the point is evidence routing. The paper combines weighted quality score, production eligibility gates, cost-adjusted performance, and Total Cost of Agent Ownership. It reports a concrete lesson: per-interaction cheapness can conflict with quality gates, and lifecycle evaluation cost remains small at the studied moderate volume. In the authors' example, continuous evaluation adds less than 11 percent overhead even for the cheapest model.
The governance lesson is not that the proposed thresholds are universal. It is that the threshold, scorer, cost formula, lifecycle state, and human sign-off should be explicit enough to audit. Otherwise "best model" means whatever the last spreadsheet or vendor dashboard made easiest to defend.
Limits and Receipts
The authors name several limits: the proof of concept is small, approval can add governance latency, evaluation cost scales linearly with agent count, MCP-native agent-to-agent patterns are still evolving, and production systems may require richer capability negotiation and scheduling optimization. Those limits are useful because they keep the paper from becoming a product brochure.
A registry-governance receipt should name the agent owner, version, deployment path, tool scope, model, prompt or policy bundle, evaluation dataset, evaluator versions, gate thresholds, weighted quality score, latency and harm metrics, cost formula, state transition, staleness policy, human approver, deprecation trigger, migration plan, retirement date, discovery visibility, and preserved audit history.
The basic question is whether an enterprise can make an agent disappear from operational discovery when the evidence says it should. If the registry cannot demote and retire agents, it is not governance. It is inventory.
Sources
- Richard Kang and Vincent Wang, Registry-Governed Agent Lifecycle: Completing EDDOps with Evaluation-Driven Registration, Promotion, and Retirement on AWS AgentCore, arXiv:2607.00345 [cs.SE], submitted July 1, 2026.
- arXiv experimental HTML for Registry-Governed Agent Lifecycle, checked for the state machine, evaluation gates, proof-of-concept scope, cost framework, MCP discovery coupling, limitations, and future-work notes.
- arXiv API record for arXiv:2607.00345, checked for title, authors, submission date, subject class, and abstract metadata.
- AWS, The future of managing agents at scale: AWS Agent Registry now in preview, checked for current Agent Registry preview details, structured records, approval workflow, discovery, and lifecycle tracking claims.