LotsAgent
Agent orchestration is when one AI agent hands work to a more specialized agent at a defined trigger point — so each agent does what it does best and the chain continues.
You've been running a single AI agent for a while. It handles email triage, drafts responses, routes leads. It's working. But at some point you realize: this agent could hand off part of its work to a more specialized agent.
A lead comes in → your first agent qualifies it → a specialized sales agent drafts the outreach → a scheduling agent confirms the meeting → your original agent logs everything to the CRM. That's the pattern.
When to Use Multiple Agents
Not every workflow needs multiple agents. Single agents handle:
- One clear goal with predictable inputs
- A workflow that doesn't need specialized decision-making
- Work where the context is always similar
Multiple agents make sense when:
- The workflow has distinct phases that need different expertise
- Different decisions require different data or criteria
- You want to scale one part without scaling the whole workflow
Example: Customer onboarding
Single agent approach: One agent handles everything — welcome email, setup instructions, feature introductions, check-in at day 7, upgrade pitch at day 14.
Problem: That's a lot for one agent to track. The setup instructions need product knowledge. The upgrade pitch needs sales context. The day 7 check-in needs customer health data.
Multi-agent approach:
- Onboarding agent — sends welcome email, setup guide, first-week tips
- Product specialist agent — answers feature questions, sends relevant tutorials
- Health agent — monitors usage, flags at-risk accounts, triggers re-engagement
- Sales handoff agent — identifies upsell candidates, drafts upgrade pitch
Each agent does one thing well. They pass work between them at defined checkpoints.
The Handoff Model
Here's how orchestration actually works in practice.
1. Define the trigger
Agent A finishes a step and creates a condition that Agent B is watching for.
Example: The lead qualification agent scores a lead above 70 → hands off to the outreach agent.
2. Pass context
Agent A passes the relevant context to Agent B. Not everything — just what Agent B needs to continue the work.
Example: Lead score, company size, industry, what the initial inquiry was about.
3. Agent B continues
Agent B picks up from where Agent A stopped. It has context, it has a goal, it executes.
4. Optional: route back
Agent B might complete a step that triggers another handoff. The chain can continue as long as each handoff is intentional.
In LotsAgent, subagents are built for this. You define the agents, the handoff triggers, and what context passes between them. Each execution is checkpointed via Inngest — if something fails, the workflow resumes from where it left off, not from scratch.
Practical Orchestration Patterns
Pattern 1: Triage → Specialist
Triage agent reads inbound requests, categorizes by type, routes to the right specialist.
- Legal question → legal specialist agent
- Technical issue → support specialist agent
- Sales inquiry → sales specialist agent
Each specialist agent has deep knowledge in its area. They don't waste time on triage — they just receive work that's already sorted.
Pattern 2: Research → Action
Research agent gathers context: company data, contact history, relevant news.
Action agent uses that context to execute: draft an email, create a task, update a record.
The research agent never sends anything. The action agent never does its own research. They hand off at a defined point.
Pattern 3: Monitor → Alert → Resolve
Monitor agent watches for conditions: a deal stalling, an account going dark, a metric crossing a threshold.
Alert agent formats and sends the notification: Slack message, Telegram alert, email.
Resolution agent takes the first response step: sends a re-engagement sequence, creates a follow-up task, flags for human review.
The monitor agent runs continuously. The other agents only activate when triggered.
Why Not Just One Smarter Agent?
You could build one agent that does everything. But:
-
Specialization — An agent that's great at triage isn't necessarily great at sales outreach. Different skills, different training data, different context requirements.
-
Scaling — If only 10% of your leads need sales follow-up, you can scale the sales agent without scaling the triage agent.
-
Maintainability — A complex single agent is harder to debug and update. When something breaks, you don't know if it's the triage logic, the routing rules, or the outreach logic.
-
Ownership — Different agents can have different owners. Marketing owns the nurture agents. Sales owns the outreach agents. Support owns the triage agents. This matters when agents need a real identity — you need to know who owns each one and who reviews its outputs.
How Agents Actually Communicate
Multi-agent systems need a way for agents to talk to each other. Two approaches exist:
Agent-to-Agent (A2A) protocols define how agents discover each other's capabilities and pass work between them. This is the open standard approach — agents from different systems can communicate if they share the protocol.
Platform-specific handoffs (like LotsAgent's subagents) use internal mechanisms to pass context and trigger the next agent in the chain. The advantage: it's built in, it works out of the box, and durable execution handles retries and checkpointing automatically.
What Orchestration Requires
Agent orchestration isn't free. It requires:
Defined handoff points — You need to know where one agent stops and another starts. Vague handoffs lead to redundant work or dropped context.
Context passing — The receiving agent needs enough context to continue without re-doing what the previous agent did.
Clear ownership — Who owns each agent? Who reviews its outputs? Who fixes it when it breaks?
Monitoring across agents — When something goes wrong, you need to trace the execution across multiple agents to find where the problem started.
LotsAgent handles this with subagent support: you define the agents, the triggers, and the handoff logic, and the platform manages the execution. Every action is logged, every handoff is visible, and you can trace any execution back to see exactly what happened.
The Practical Setup
On LotsAgent, here's what orchestration looks like:
Agent A: Lead qualification
- Reads inbound inquiry
- Checks CRM for existing contact
- Researches company
- Scores lead based on your criteria
- Hands off to Agent B if score > 70
Agent B: Sales outreach
- Receives lead score, company data, inquiry content
- Drafts personalized outreach email
- Routes to sales rep in Slack
Agent C: Follow-up scheduling
- Monitors for reply to outreach
- If no reply in 48 hours, sends follow-up
- If reply received, routes to sales rep for meeting scheduling
Each agent has a clear role. Each handoff is defined. You monitor the whole workflow from one place. Agents get access to 100+ tools via Composio without needing custom integrations for each one.
When to Start Small
You don't need to build a complex multi-agent system on day one. Start with two agents:
- A primary agent that does most of the work
- A single subagent that handles one specific task the primary agent struggles with
Example: Your inbox agent handles triage and response drafting, but it doesn't know your product well enough for technical questions. Add a technical specialist subagent that handles those handoffs.
Expand from there as you find more natural handoff points.
Create your first agent free at lotsagent.com.
FAQ: Agent Orchestration
What's the difference between agent orchestration and just running multiple single agents?
Single agents work independently — each one has a goal and executes toward it. Orchestration connects agents so they hand off work to each other at defined points. The output of Agent A becomes the input of Agent B. The chain continues as needed. Multiple single agents without handoffs is just parallel execution, not orchestration.
When should I split work across multiple agents?
Split work when: different phases of the workflow need different expertise, you want to scale one part independently, or the workflow is complex enough that one agent doing everything becomes hard to maintain and debug. Also consider splitting when different agents need access to different tools or data sources — a specialist agent can have deep tool access without exposing it to the whole workflow.
How do agents know when to hand off work?
You define the handoff triggers. In LotsAgent, subagents watch for conditions created by the parent agent. When the trigger fires, the next agent activates with the relevant context. The trigger can be a score threshold, a status change, a time delay, or any measurable condition.
What's the simplest multi-agent workflow to start with?
Two agents: a triage/qualification agent and a specialist agent. The first agent reads inbound work and routes to the second based on type or score. The second agent handles the specialized work. Start there. Add more agents only when you find a natural handoff point that the current setup can't handle cleanly.
Can agents hand off work back to the original agent?
Yes. The orchestration chain can be bidirectional. An agent might complete a step, hand off to another agent, and receive the result back with additional context before continuing. The chain doesn't have to be linear — it can loop back, split, or merge depending on what the workflow requires.
What happens when something breaks in an orchestration chain?
Execution logs show you which agent failed, at what step, and what it was trying to do. You can trace the problem across the chain and fix the specific agent without rebuilding the entire workflow. Durable execution means the workflow can resume from the last successful checkpoint — you don't lose all the progress when something goes wrong.