Swarm Management
Human-in-the-Loop
Human-in-the-Loop
Human-in-the-Loop (HITL) is one of HIVE Protocol's most powerful features. It allows you to maintain the perfect balance between AI automation and human oversight. Rather than choosing between fully autonomous AI or manually doing everything yourself, HITL gives you granular control over how and when you participate in agent workflows.
Why Human-in-the-Loop Matters
AI agents are incredibly capable, but there are situations where human judgment is essential:
┌─────────────────────────────────────────────────────────────────┐
│ The Automation Spectrum │
├─────────────────────────────────────────────────────────────────┤
│ │
│ Full Manual Collaborative Supervised Autonomous │
│ │ │ │ │ │
│ ●───────────────●──────────────────●───────────────● │
│ │ │ │ │ │
│ You do You guide You review Agents │
│ everything agents outputs handle it │
│ │
│ ▲ │
│ │ │
│ HIVE Protocol HITL │
│ gives you control │
│ across this spectrum │
│ │
└─────────────────────────────────────────────────────────────────┘Key benefits of HITL:
| Benefit | Description |
|---|---|
| Quality Control | Review and approve critical outputs before they're used |
| Domain Expertise | Inject specialized knowledge agents might not have |
| Risk Mitigation | Catch and correct errors before they compound |
| Learning Opportunity | Understand how agents work to improve your prompts |
| Compliance | Maintain audit trails and human oversight for regulated tasks |
The Three Collaboration Modes
HIVE Protocol provides three distinct collaboration modes, each designed for different scenarios and levels of involvement.
Mode 1: Observe Mode
Observe mode lets you watch your agent swarm work without interfering. You're a silent observer, monitoring the conversation and outputs without participating.
┌─────────────────────────────────────────────────────────────────┐
│ OBSERVE MODE [* . .] │
├─────────────────────────────────────────────────────────────────┤
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ Agent: Researcher 10:23 AM │ │
│ │ I've analyzed the market data. Key findings: │ │
│ │ - Market growing at 12% annually │ │
│ │ - Three main competitors identified │ │
│ │ - Customer preference shifting to mobile │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ Agent: Analyst 10:24 AM │ │
│ │ Thanks for that data. Let me cross-reference with │ │
│ │ our internal metrics. The mobile shift aligns with │ │
│ │ what we're seeing in our analytics... │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ You are observing │ │
│ │ Agents are working autonomously │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────┘When to use Observe Mode:
| Scenario | Why Observe Works |
|---|---|
| Routine tasks | Agents handle repetitive work while you focus elsewhere |
| Learning | Understand how agents collaborate and reason |
| Quality sampling | Periodically review outputs without interrupting flow |
| High-trust workflows | Established processes that agents execute reliably |
| Background research | Let agents gather information while you do other work |
Observe Mode settings:
observe_mode:
notifications:
on_completion: true
on_error: true
on_milestone: false
auto_review:
enabled: true
priority_threshold: "high"
intervention_triggers:
error_count: 3
confidence_threshold: 0.7
timeout_minutes: 30Mode 2: Collaborate Mode
Collaborate mode creates a true partnership between you and your agents. You can offer suggestions, share ideas, and provide guidance, while agents build upon your input with their own analysis and creativity.
┌─────────────────────────────────────────────────────────────────┐
│ COLLABORATE MODE [. * .] │
├─────────────────────────────────────────────────────────────────┤
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ You 2:15 PM │ │
│ │ I've been thinking about the pricing strategy. │ │
│ │ What if we offered a freemium tier to capture │ │
│ │ more of the SMB market? │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ Agent: Market Analyst 2:15 PM │ │
│ │ That's an interesting angle! Building on that idea, │ │
│ │ I'd suggest we analyze the freemium models of │ │
│ │ three successful competitors: Slack, Notion, and │ │
│ │ Figma. Each took a different approach... │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ You are collaborating │ │
│ │ Share ideas and agents will build upon them │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────┘When to use Collaborate Mode:
| Scenario | Why Collaborate Works |
|---|---|
| Brainstorming | Agents expand on your ideas with additional perspectives |
| Strategic planning | Combine human intuition with data-driven analysis |
| Creative projects | Get AI inspiration while maintaining creative control |
| Problem solving | Work through complex issues together |
| Learning new domains | Agents fill knowledge gaps while you guide direction |
Effective collaboration techniques:
TECHNIQUE 1: Seed and Expand
You: "I'm thinking about sustainability as a brand differentiator."
Agent: [Expands with market trends, competitor analysis, strategies]
TECHNIQUE 2: Challenge and Refine
You: "What's wrong with this approach?"
Agent: [Provides critical analysis, then suggests improvements]
TECHNIQUE 3: Parallel Exploration
You: "Let's explore three different directions simultaneously."
Agents: [Each agent develops a different direction]
TECHNIQUE 4: Build on Build
You: "Great point about mobile. What about tablet users?"
Agent: [Extends the analysis to cover your new angle]Mode 3: Direct Mode
Direct mode gives you precise control over your agents. Your instructions are followed exactly as given, making this mode ideal for tasks requiring specific outputs or when you know exactly what you need.
┌─────────────────────────────────────────────────────────────────┐
│ DIRECT MODE [. . *] │
├─────────────────────────────────────────────────────────────────┤
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ You (Command) 4:30 PM │ │
│ │ Write a product announcement following this exact │ │
│ │ structure: │ │
│ │ 1. Hook (1 sentence, max 15 words) │ │
│ │ 2. Problem statement (2 sentences) │ │
│ │ 3. Solution (3 bullet points) │ │
│ │ 4. CTA (1 sentence with link placeholder) │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ Agent: Copywriter 4:30 PM │ │
│ │ Following your exact structure: │ │
│ │ │ │
│ │ 1. HOOK: Your team meetings are about to get 10x │ │
│ │ more productive. │ │
│ │ │ │
│ │ 2. PROBLEM: Teams waste an average of 31 hours per │ │
│ │ month in unproductive meetings. │ │
│ │ │ │
│ │ 3. SOLUTION: │ │
│ │ - AI-generated agendas based on project context │ │
│ │ - Real-time action item capture │ │
│ │ - Automated follow-up reminders │ │
│ │ │ │
│ │ 4. CTA: Start running better meetings today - │ │
│ │ [LINK: Try free for 14 days] │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ You are directing │ │
│ │ Agents follow your instructions precisely │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────┘When to use Direct Mode:
| Scenario | Why Direct Works |
|---|---|
| Specific formatting | Output must match exact specifications |
| Corrections | Fix mistakes precisely without interpretation |
| Compliance requirements | Specific language or structure is mandatory |
| Quality control | Enforce standards and catch deviations |
| Template filling | Populate predefined structures accurately |
Command syntax for Direct Mode:
EXPLICIT INSTRUCTIONS
- "Write exactly 3 paragraphs, each 2-3 sentences."
- "List precisely 5 benefits, numbered 1-5."
- "Format as a markdown table with columns: Name, Date, Status"
CONSTRAINT COMMANDS
- "Do not include any pricing information."
- "Keep the total response under 200 words."
- "Avoid these terms: [list of terms to avoid]"
STRUCTURE COMMANDS
- "Follow this exact template: [template]"
- "Mirror the structure of this example: [example]"
- "Output as JSON with this schema: [schema]"Switching Between Modes
One of HITL's most powerful features is the ability to switch modes mid-conversation.
Mode Switching Interface
┌─────────────────────────────────────────────────────────────────┐
│ Mode Selector │
├─────────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ OBSERVE │ │ COLLABORATE │ │ DIRECT │ │
│ │ │ │ [Active] │ │ │ │
│ └─────────────┘ └─────────────┘ └─────────────┘ │
│ │
│ Click any mode to switch. Agents will be notified. │
│ │
└─────────────────────────────────────────────────────────────────┘Smooth Transitions
SCENARIO: Research task that needs refinement
Step 1: Start in OBSERVE
Let agents gather initial research autonomously
Agent A: "Found 15 relevant studies..."
Agent B: "Cross-referencing with industry reports..."
[You watch silently]
Step 2: Switch to COLLABORATE
Join the discussion to guide direction
You: "The healthcare angle is most relevant for us."
Agent A: "Great insight! Let me focus on healthcare..."
Step 3: Switch to DIRECT
Get specific outputs for the final deliverable
You: "Summarize the top 5 healthcare findings in a table"
Agent B: [Produces exact table format requested]Mode Transition Best Practices
| From | To | When | Example Trigger |
|---|---|---|---|
| Observe | Collaborate | Agents going off-track | "Y is more relevant than X" |
| Observe | Direct | Need immediate output | "Give me a summary now" |
| Collaborate | Observe | Discussion is productive | "I'll let you continue" |
| Collaborate | Direct | Need to lock in decisions | "Use approach B exactly" |
| Direct | Collaborate | Need creative exploration | "Let's brainstorm improvements" |
| Direct | Observe | Task is well-defined | "Continue with this format" |
Intervention Strategies
Even in Observe or Collaborate modes, there are times when you need to intervene.
Strategy 1: Gentle Steering
Guide the conversation without disrupting flow:
Agent A: "Let's look at the technical architecture first..."
Agent B: "Good idea. Starting with the database layer..."
You: "Just a thought - might be worth considering the user
experience implications as you go. No rush."
Agent A: "Great point! Let me incorporate UX considerations
into our technical discussion..."When to use: Agents are productive but missing an angle you care about.
Strategy 2: Context Injection
Provide critical information agents don't have:
Agent: "Based on typical enterprise sales cycles, I'd
estimate 6-9 months to close..."
You: "Important context: We already have a champion at the
prospect company who's pushing for a Q1 decision.
Also, they have budget that expires in March."
Agent: "That changes things significantly! Let me revise
my approach to capitalize on this urgency..."When to use: You have insider knowledge, recent developments, or constraints.
Strategy 3: Error Correction
Fix mistakes immediately before they propagate:
Agent: "Since you're using Python 2.7, the syntax would be..."
You: "Quick correction: We're on Python 3.11, not 2.7."
Agent: "Thanks for catching that! Python 3.11 changes things.
Let me update my recommendations..."When to use: Agent is working from incorrect assumptions.
Strategy 4: Scope Management
Keep the conversation focused and productive:
Agent A: "We should also consider blockchain integration..."
Agent B: "Yes! And maybe IoT connectivity..."
You: "Let's stay focused on the core MVP. We can explore
those advanced features in a future phase."
Agent A: "You're right. Focusing on MVP essentials:
1. User authentication
2. Core data management
3. Basic reporting..."When to use: Discussion is expanding beyond useful scope.
Strategy 5: Quality Feedback
Help agents improve their outputs iteratively:
Agent: [Produces a draft marketing email]
You: "Good start! A few refinements:
- The opening feels too generic, make it more personal
- The CTA is buried, move it higher
- Add a specific deadline to create urgency"
Agent: "Great feedback! Here's the revised version..."When to use: Output is close but needs refinement.
Approval Workflows
For high-stakes tasks, HIVE Protocol supports structured approval workflows.
Workflow Types
TYPE 1: Simple Approval
Agent: "I've drafted the customer response. Ready to send?"
[Preview of response]
[Approve] [Edit] [Reject]
TYPE 2: Checkpoint Workflow
┌─────────────────────────────────────────────────────────────────┐
│ Content Production Workflow │
├─────────────────────────────────────────────────────────────────┤
│ Step 1: Topic Research [Done] │
│ Step 2: Outline Creation [Awaiting Review] │
│ Step 3: First Draft [Pending] │
│ Step 4: Final Review [Approval Required] │
│ Step 5: Publish [Approval Required] │
└─────────────────────────────────────────────────────────────────┘
TYPE 3: Parallel Approval
Multiple stakeholders review simultaneously:
Legal Review: Approved
Finance Review: Pending
Executive Review: Pending
Status: 1 of 3 approvals receivedConfiguring Approval Requirements
approval_workflow:
name: "Content Publication"
checkpoints:
- step: "outline"
type: "review"
required: false
notification: true
- step: "final_draft"
type: "approval"
required: true
approvers: ["content_lead"]
timeout_hours: 24
- step: "publish"
type: "approval"
required: true
approvers: ["content_lead", "legal"]
require_all: false
on_rejection:
action: "return_to_previous"
notify: true
require_feedback: trueApproval Actions
| Action | Effect | When to Use |
|---|---|---|
| Approve | Workflow continues | Output meets requirements |
| Approve with Comments | Continue with notes | Minor improvements possible |
| Request Changes | Return for revision | Specific issues need addressing |
| Reject | Stop the workflow | Fundamental problems |
| Delegate | Forward to another | Someone else is better positioned |
| Defer | Pause workflow | Need more time or information |
Advanced HITL Patterns
Pattern 1: Progressive Autonomy
Start with high oversight, then reduce as confidence builds:
WEEK 1: Direct Mode - Explicit instructions, review every output
WEEK 2: Collaborate Mode - Guide direction, review key outputs
WEEK 3: Observe Mode with Checkpoints - Review at milestones
WEEK 4+: Full Observe Mode - Periodic quality sampling onlyPattern 2: Exception-Based Intervention
Configure rules for automatic alerts:
intervention_rules:
confidence_alert:
trigger: "agent_confidence < 0.7"
action: "notify_and_pause"
disagreement_alert:
trigger: "agents_disagree"
action: "notify"
high_stakes_alert:
trigger: "output_type in ['contract', 'public_statement']"
action: "require_approval"Pattern 3: Role-Based Involvement
Different team members participate in different ways:
Executive: OBSERVE mode - Periodic dashboards, final approvals
Manager: COLLABORATE mode - Set direction, guide decisions
Specialist: DIRECT mode - Domain expertise, precise executionMeasuring HITL Effectiveness
Track these metrics to optimize your processes:
| Metric | What It Measures | Target |
|---|---|---|
| Intervention Rate | How often humans step in | Decreasing over time |
| Approval Time | How long approvals take | Consistent, not blocking |
| Rejection Rate | How often outputs rejected | Low (< 10%) |
| Mode Distribution | Time in each mode | Shifting toward Observe |
Best Practices Summary
1. Choose the Right Starting Mode
| Task Type | Recommended Start |
|---|---|
| Exploration/Research | Observe |
| Strategic Planning | Collaborate |
| Precise Execution | Direct |
| New Workflow | Direct |
| Proven Workflow | Observe |
2. Communicate Clearly
GOOD: "Focus on the mobile user experience,
specifically the checkout flow."
BAD: "Make it better for users."3. Trust But Verify
- Set clear objectives upfront
- Review outputs at natural breakpoints
- Provide feedback that improves future performance
- Gradually reduce oversight as confidence grows
4. Match Detail to Mode
| Mode | Your Input Style |
|---|---|
| Observe | Silent or minimal prompts |
| Collaborate | Ideas, questions, suggestions |
| Direct | Explicit, detailed instructions |
Related Documentation
- [Creating Swarms](/docs/swarms/creating-swarms): Set up swarms to use with HITL
- [Communication Patterns](/docs/swarms/communication-patterns): How agents communicate
- [Context Blocks](/docs/swarms/context-blocks): Provide context to guide behavior