Skip to Content
DocsOntologyOntology for Agents

Ontology for Agents

The Runtime ontology is a machine-readable vocabulary for scoring, feedback, campaigns, and plugins. AI agents should look up concepts and validate payloads before they call invoke / response. That keeps tools aligned with POST /invocations then POST /response (singular), and with params as a JSON string.

The Runtime Dashboard MCP tab includes an Ontology Explorer that calls the same REST catalog the Java runtime exposes.

Runtime Ontology Explorer — search concepts and related MCP tools

Try the hosted runtime at https://medemo.ecosystem.ai  (point the dashboard Runtime URL at that host), or run locally (default 8091, local debug often 8099).

Why agents use the ontology

ProblemOntology answer
Invented field names (/responses, nested params)Concept cards for InvocationRequest and ResponseFeedback
Wrong closed-loop orderRecipe GET /ontology/runtime/recipes/closed-loop
Payload drift vs pluginsvalidateInvokeRequest / validateResponseRequest before scoring
Which MCP tool to callEach concept lists related tools (invoke, listCampaigns, …)

Look up concepts through the runtime catalog below. The documentation site does not serve Turtle files. See Runtime Closed Loop.

REST catalog

Same content as MCP resources. Prefix {RUNTIME} (for example http://localhost:8091).

MethodPathPurpose
GET/ontology/runtimeModule summary
GET/ontology/runtime/concepts?q=invocationSearch by label or local name
GET/ontology/runtime/concepts/{localName}One concept card
GET/ontology/runtime/recipes/closed-loopClosed-loop recipe JSON
POST/ontology/runtime/validate/invokePre-flight invoke body
POST/ontology/runtime/validate/responsePre-flight response body
curl "http://localhost:8091/ontology/runtime/concepts?q=invocation&limit=20"

MCP tools and resources

On Runtime 0.9.7+, POST {RUNTIME}/mcp exposes:

ToolUse
lookupRuntimeConceptSearch concepts (same as the Explorer query box)
getClosedLoopRecipeCanonical listCampaigns → validate → invoke → validate → response
validateInvokeRequestRequire params as a JSON string
validateResponseRequestRequire uuid + offer_name
invoke / responseScore and learn

Resources: ontology://runtime, ontology://runtime/closed-loop, ontology://runtime/examples, ontology://runtime/concept/{localName}.

Recommended flow:

  1. getClosedLoopRecipe or read ontology://runtime/closed-loop
  2. listCampaigns
  3. validateInvokeRequest → invoke (save uuid and offer_name)
  4. validateResponseRequest → response

Protocol detail: MCP Support. Workbench Operator agents use a different MCP server (POST Workbench :8001/mcp) for deployments and enrichment — do not mix the two /mcp surfaces.

Workbench project kind

The entities ontology tells a Workbench agent whether a project is dynamic, static, or hybrid. Derive engagementKind from hasDynamicEngagement, hasPredictor, hasFeatureStoreFrame, and deployment scoringMode. Do not classify from the stored projectType string.

learningAlgorithm is only the Dynamic Engagement dropdown: epsilonGreedy, binaryThompson, naiveBayes, QLearning, behaviorAlgos, Network. When the approach is behaviorAlgos, behavioralType is one of lossAversion, riskAversion, prospectTheory, sentimentalEquilibrium, coverageAwareThompson, longTailBoostMF, generative.

Spend Personality and Money Personality are lookup-only (scoringMode=lookup). Any installation lookup is valid when the document has the fields the post-score plugin copies. A personality plugin stays lookup even if is_multi_armed_bandit is set. On spend_personality, spend_personality and money_personality are lookup deployments and dynamic_recommender_behavior is dynamic engagement.

Last updated on