Skip to content
ConversionIQ Docs

Agent-Orchestrated Execution

Purpose

Section titled “Purpose”

This document defines the future runtime workflow for handling user input through a chain of specialized runtime agents.

This is a product/runtime workflow, not the engineering delivery workflow documented in docs/ai-governance/**.

Related docs:

Core flow

Section titled “Core flow”
flowchart TD
  userInput[UserInput]
  interactionEvent[InteractionEvent]
  intentAgent[IntentAgent]
  policyAgent[PolicyAgent]
  plannerAgent[PlannerAgent]
  executionAgent[ExecutionAgent]
  validationAgent[ValidationAgent]
  learningCapture[LearningCapture]
  userOutcome[UserOutcome]


  userInput --> interactionEvent
  interactionEvent --> intentAgent
  intentAgent --> policyAgent
  policyAgent --> plannerAgent
  plannerAgent --> executionAgent
  executionAgent --> validationAgent
  validationAgent --> learningCapture
  learningCapture --> userOutcome

Workflow stages

Section titled “Workflow stages”

1) Input normalization

Section titled “1) Input normalization”

Convert voice, text, or UI interaction into a normalized InteractionEvent.

2) Intent interpretation

Section titled “2) Intent interpretation”

The IntentAgent determines:

  • what the user is trying to achieve
  • which entities are involved
  • whether clarification is required

3) Policy and scope check

Section titled “3) Policy and scope check”

The PolicyAgent determines:

  • whether the action is permitted
  • whether tenant/workspace scope is valid
  • whether approval is required
  • whether the request is safe to continue

4) Planning

Section titled “4) Planning”

The PlannerAgent determines:

  • whether the task is feasible
  • which workflow should be used
  • which tools or specialized agents are needed
  • whether the task should be split into multiple steps

5) Execution

Section titled “5) Execution”

The ExecutionAgent performs:

  • retrieval
  • tool or API calls
  • state updates
  • response generation
  • recommendation generation

6) Validation

Section titled “6) Validation”

The ValidationAgent verifies:

  • whether the goal was actually achieved
  • whether the result is correct
  • whether policy and safety constraints were preserved

7) Learning capture

Section titled “7) Learning capture”

The system records:

  • what was attempted
  • what succeeded or failed
  • what required clarification
  • whether the user accepted the outcome

Clarification and escalation branches

Section titled “Clarification and escalation branches”

The happy path must not be the only documented path.

The runtime must be able to:

  • request clarification before execution
  • stop on low confidence or blocked policy
  • ask for approval before sensitive actions
  • hand off to a human when automation is unsafe or insufficient
flowchart TD
  intent[IntentAssessment]
  clarify[ClarifyWithUser]
  approval[RequestApproval]
  execute[Execute]
  handoff[HumanHandoff]


  intent --> clarify
  intent --> approval
  intent --> execute
  approval --> execute
  approval --> handoff
  clarify --> execute
  clarify --> handoff

Runtime agent boundaries

Section titled “Runtime agent boundaries”

Recommended runtime roles:

  • IntentAgent
  • PlannerAgent
  • PolicyAgent
  • ExecutionAgent
  • ValidationAgent
  • LearningAgent
  • RecommendationAgent

Boundary rule:

  • runtime agents serve product behavior
  • delivery agents serve engineering workflow

Do not reuse Neo, Coder, QA, or other delivery-agent labels as runtime behavior classes unless the product intentionally wants that overlap documented and governed.

Trust and audit requirements

Section titled “Trust and audit requirements”

This workflow must preserve:

  • workspace and tenant isolation
  • approval gates for sensitive actions
  • explicit auditability for meaningful state changes
  • explainability of why a decision path was chosen

Reference:

MVP-compatible path

Section titled “MVP-compatible path”

Start with:

  • text and UI interactions
  • a small planner/executor/validator loop
  • recommendation capture
  • human approval for risky actions

Add later:

  • voice as another input modality
  • richer memory
  • adaptive recommendations
  • more specialized runtime agents