Skip to main content

TAP Agents

Overview​

TAP agents use the Unified Agent Platform for profile loading, model resolution, and prompt workbench support.

The active TAP product agent set is intentionally small:

  • tap-caption-tagging-agent
  • tap-query-agent
  • tap-topic-detection-agent

The shared docs assistant still exists in the platform docs-chat layer, but it is not part of the TAP domain runtime. Legacy TAP chat, image-caption, content-tagger, content-safety blocklist, and recovered agents are retired from the TAP runtime and should not appear as current TAP product agents.

TAP Caption And Tagging Agent​

  • Registry ID: tap-caption-tagging-agent
  • Profile ID: tap-caption-tagging-agent
  • Purpose: Generate tag and caption suggestions from caller-provided post context and media.
  • Tools: None.
  • Writes to TAP: No.
  • Code: packages/domain-tap/src/agents/tap-caption-tagging-agent.ts

Runtime requests pass options.maxTagsPerPost, which is included in the per-post user prompt. The service also normalizes returned tag names to Title Case before responding.

Response shape:

{
  "caption": "Students celebrate graduation on campus.",
  "tags": ["Graduation", "Campus", "Students"],
  "confidence": "high",
  "reasoning": "The image and post text indicate a graduation scene."
}

TAP Query Agent​

  • Registry ID: tap-query-agent
  • Profile ID: tap-query-agent
  • Purpose: Synthesize structured TAP query responses from retrieved context.
  • Code: packages/domain-tap/src/agents/tap-query-agent.ts

TAP Topic Detection Agent​

  • Registry ID: tap-topic-detection-agent
  • Profile ID: tap-topic-detection-agent
  • Purpose: Classify completed or quiet TAP student conversations into stable topic taxonomy keys.
  • Tools: None.
  • Writes to TAP: No.
  • Code: packages/domain-tap/src/agents/tap-topic-detection-agent.ts

The topic detection service passes ordered conversation context, local candidate topic rankings, provider lane, and analysis mode into the agent. fast mode sends a compact candidate prompt; full mode sends the full taxonomy plus candidates. Both modes require the same response shape with topics, messageFindings, and fallback so auditability is preserved.

The seeded profile defaults to the gpt-5.4-nano model with reasoningEffort: "minimal" and maxOutputTokens: 5000. Target environments must have a matching enabled model capability for this override; otherwise the agent factory falls back to the environment default model.

Removed Legacy Agents​

The old standalone image caption agents, TAP chat agent, TAP content tagger, TAP content safety profiles, Azure blocklist agent, and recovered agent are retired for the TAP product runtime. The current tagging integration uses the combined caption and tagging agent above.