The Agent Registry: DNS for the Intelligence Layer

A New Coordination Problem

As the number of AI agents in an organisation grows from one to hundreds, a new problem emerges: how do you know which agent to call for a given task?

At 100 agents, maintaining manual routing tables is a full-time job. At 1,000 agents, it becomes impossible. This is the problem the Agent Registry solves.

DNS as a Mental Model

The Domain Name System solves a simple problem elegantly: given a human-readable name (example.com), return a machine-readable address (93.184.216.34).

An Agent Registry does the same one level up: given a task description, return the best agent to handle it.

The Composite Quality Score

The registry returns agents ranked by a composite score:

Composite = (Accuracy × 0.30) + (Reliability × 0.25) +
            (Latency × 0.20) + (Cost × 0.15) +
            (Popularity × 0.10)

Accuracy (30%) — successful execution percentage in the last 30 days.

Reliability (25%) — success rate squared, to strongly penalise frequent failures.

Latency (20%) — compared to the category median. Faster = higher score.

Cost (15%) — lower cost per call scores higher.

Popularity (10%) — log-scaled total execution count plus rating signal.

Capability Tags: Structured Discovery

Beyond scores, the registry indexes agents by capability tags — structured, machine-readable strings:

text_summarisation
legal_document_analysis
sentiment_classification
entity_extraction
code_generation:python
structured_data_extraction

A registry query can filter by capability before ranking by score:

GET /api/registry/search?capabilities=legal_document_analysis,structured_data_extraction

Querying the Registry from Code

const response = await fetch(
  '/api/registry/search?q=summarise+legal+document&limit=3',
  { headers: { 'Authorization': `Bearer ${apiKey}` } }
)
const { agents } = await response.json()

// agents[0] is the highest-scoring match
const bestAgent = agents[0]
console.log(bestAgent.name, bestAgent.composite_score)