Agent Registry

"Before deploying a new agent, know what agents already exist."

Context

An organization has multiple agent deployments — customer support agents, code generation pipelines, monitoring agents, data analysis agents. Teams deploy new agents independently. There is no central view of what agents exist, what archetype each follows, what spec governs each, or who owns each.

Problem

Without a registry, the organization cannot answer basic questions: How many agents do we have? Which ones have write access to production systems? Which ones were last reviewed more than six months ago? Duplicate agents are deployed for the same function. Governance reviews cannot enumerate the systems under governance.

Concrete scenario: A mid-size fintech company. Engineering deploys "MarketAnalysis-v1" (autonomy archetype, read-only). Two months later, the Data team deploys "MarketWatch" (autonomous archetype, write access to recommendations database) to run the same function. Neither team knows about the other. Six months goes by with both running in production. An audit finds two agents doing the same job with different governance oversight. The Data team's agent has never been reviewed. The Engineering team doesn't know MarketWatch exists. The company cannot answer: "How many agents have production write access?"

Forces

  • Visibility vs. maintenance. A registry only earns its value if it's kept current. The registry must be wired into the deployment pipeline, not maintained as a separate spreadsheet, or it goes stale within months.
  • Governance compliance vs. deployment friction. Heavy registry requirements at deploy time slow teams down. Light requirements produce a registry whose data is unreliable. The discipline is making the minimum required fields trivial to populate (auto-derived from spec) and the optional ones defensible.

The Solution

Maintain a discoverable agent registry — a catalog of all deployed agent systems with their governance metadata.

Registry entry structure:

  • Agent name and identifier
  • Archetype classification
  • Current spec version (link to spec document)
  • Owner (team and individual)
  • Deployment status (active, staging, deprecated)
  • Last governance review date
  • Tool manifest summary (effect classes in use)
  • Escalation path

The registry is the organizational view of the agent fleet. It is maintained alongside the archetype catalog and updated when agents are deployed, modified, or retired.

Example: The fintech company implements a registry at governance.internal/agents/:

{
  "agents": [
    {
      "id": "market-analysis-v1",
      "name": "Market Analysis (Engineering)",
      "archetype": "autonomy:read_only",
      "spec_version": "market-analysis/v2.3.1",
      "owner": {"team": "Engineering", "contact": "eng-leads@company.com"},
      "status": "active",
      "last_review": "2026-02-15",
      "tools_used": ["data:read", "api:market_data"],
      "escalation": "eng-sre@company.com"
    },
    {
      "id": "market-watch",
      "name": "Market Watch (Data Team)",
      "archetype": "autonomy:read_write",
      "spec_version": "market-watch/v1.0.0",
      "owner": {"team": "Data", "contact": "data-leads@company.com"},
      "status": "active",
      "last_review": null,
      "tools_used": ["data:read", "database:write", "api:market_data"],
      "escalation": "data-sre@company.com"
    }
  ]
}

Query: "Which agents have database write access?" → Returns market-watch. "Which agents haven't been reviewed in 6 months?" → Returns market-watch. Governance team schedules a review. When a new agent is proposed, the team checks registry first: "Market analysis already exists — propose consolidation or a different function."

Resulting Context

  • Duplication is visible and preventable — teams can see what already exists before deploying
  • Governance is fleet-wide — compliance checks can query all agents systematically
  • Ownership is clear — every agent has a designated owner and escalation path
  • Risk is quantified — how many write-access agents? How many overdue for review?

Therefore

Maintain a registry of all deployed agents with their archetype, spec version, owner, and governance status. The registry enables fleet-level governance, prevents duplication, and makes the organization's agent landscape visible.

Connections