ClassifierAgent
Purpose: Route intent to the right workflow.
Inputs: user_request
Outputs: intent_label
Failure modes: ambiguous labels
Compose with: Router → Planner
Agent Catalog
Reusable micro-agents
Each agent has a purpose, typed inputs/outputs, known failure modes, and composition hints.
ExtractorAgent
Purpose: Structured output with strict schema.
Inputs: document_text
Outputs: json_payload
Failure modes: missing fields
Compose with: ValidatorAgent
WriterAgent
Purpose: Constrained tone and format.
Inputs: outline, tone
Outputs: formatted_copy
Failure modes: tone drift
Compose with: PolicyAgent
SummarizerAgent
Purpose: Turn long docs into concise briefs.
Inputs: long_text
Outputs: summary
Failure modes: missing key points
Compose with: WriterAgent
ResearchAgent
Purpose: Web research gated by tool policy.
Inputs: question, sources
Outputs: notes
Failure modes: stale sources
Compose with: QA/EvalAgent
QA/EvalAgent
Purpose: Self-check with rubric.
Inputs: output, rubric
Outputs: score, notes
Failure modes: rubric mismatch
Compose with: SupervisorAgent
PolicyAgent
Purpose: Redaction + compliance enforcement.
Inputs: candidate_output
Outputs: redacted_output
Failure modes: over-redaction
Compose with: ValidatorAgent
ValidatorAgent
Purpose: Validate schema + constraints.
Inputs: schema, output
Outputs: valid, violations
Failure modes: false positives
Compose with: ExtractorAgent
SupervisorAgent
Purpose: Monitor, retry, and fallback.
Inputs: agent_status
Outputs: retry_plan
Failure modes: noisy signals
Compose with: QA/EvalAgent