KorinAI

Agent Orchestration

Orchestrate multiple agents as a team—define roles, set routing and handoffs, share context, and grant per‑role tool access so work flows reliably between agents.

Overview

Agent Orchestration lets you build specialized collaborators that work as a team:

  • Roles: e.g., Researcher, Writer, Reviewer.
  • Routing: intent/keyword rules, manual handoffs, or explicit commands.
  • Shared context: internal knowledge, variables, policies, prompt templates.
  • Tools: allow per‑role tool access (built‑in and custom tools).

Core Concepts

  • Roles and responsibilities

    • Give each role a clear mandate (input → output). Keep scopes narrow.

  • Routing and handoff

    • Automatic (intent, keywords) or manual (user/agent directive).

  • Shared context

    • Attach Notes/Files, Variables (readonly for system values), and Prompt Templates.

  • Tooling per role

    • Enable only the tools a role needs; set confirmation for side‑effects.

  • Where this lives in the UI

    • Your Agents tab: create and manage child agents that belong to your account. Configure persona/behavior, enable tools, and set defaults.
    • Agent Settings → Agent-to-Agent: define connections and workflow between agents (who can delegate to whom, routing, and handoff rules).

Getting Started

  1. Open Dashboard → Settings.
  2. Create child agents under Your Agents:
    • Click Add Agent, set name/username, configure persona/behavior.
    • Enable required tools (Built-in Agents, Custom) and confirmations.
    • Optionally attach Prompt Templates and Knowledge for grounding.
  3. Configure orchestration under Agent Settings → Agent-to-Agent:
    • Add connections: select source agent → target agent, give the connection a clear name.
    • Define when to delegate (intent/keywords) or keep it manual.
    • Tip: Use descriptive names like "Research→Write Draft" to make handoffs obvious in reviews.
  4. Add shared Variables (readonly where appropriate) and attach Knowledge sources.
  5. Save and test the workflow in Chat (try a prompt that triggers a handoff).

Route When Conditions

Both Built-in Agents and Agent-to-Agent connections support optional Route When conditions that guide the AI on when to invoke specific agents or route to specific agents.

For Built-in Agents:

  • Navigate to Agent Settings → Built-in Agents.
  • Enable an agent, expand the "Route When" section.
  • Click "Edit" to specify when this agent should be used.
  • Example: "user asks about web search or needs current information"

For Agent-to-Agent connections:

  • Navigate to Agent Settings → Agent-to-Agent.
  • Create or select an existing connection.
  • Click the chevron to expand the routing condition section.
  • Click "Edit" to specify when to route to this agent.
  • Example: "user needs detailed research or information gathering"

Benefits:

  • Intelligent routing: The AI automatically selects the right agent or tool based on user intent.
  • Reduced manual intervention: Clear routing rules minimize the need for explicit routing commands.
  • Better orchestration: Agents can work together more seamlessly with well-defined boundaries.
  • Context-aware delegation: The parent agent understands when to delegate and to whom.

Best practices for Route When conditions:

  • Use clear, natural language descriptions.
  • Focus on user intent and task characteristics.
  • Avoid overlapping conditions between agents.
  • Test your conditions with real user queries.
  • Update conditions based on actual usage patterns.

Orchestration Patterns

  • Hand‑off pipeline
    • Researcher gathers sources → Writer drafts → Reviewer polishes and cites.
  • Parallel split
    • Research + Data analysis run in parallel → Merge into one answer.
  • Guardrail review
    • Reviewer checks style, PII, policy, or compliance before send.

Tip: Use Agent Settings → Agent-to-Agent to encode these patterns explicitly as connections, and keep role prompts in Prompt Templates for consistency.

Examples

  • Release Notes Flow

    • Input: sprint notes, issue list
    • Roles: Researcher, Writer, Reviewer
    • Output: concise release notes with links and sign‑off checklist

  • Sales Brief Flow

    • Input: customer website, discovery notes
    • Roles: Researcher (web), Analyst (CSV), Writer
    • Output: 1‑pager brief with pain points and tailored value props

Best Practices

  • Keep roles narrow and well‑named; expand only as needed.
  • Start simple (2–3 roles), then iterate on routing precision.
  • Use Need confirmation for tools/roles with side effects; log agent actions.
  • Standardize tone/structure with Prompt Templates; ground with internal knowledge.

Related Topics

Previous
← Custom Tools
Next
Prompt Templates →