jean-claude-cli

name	jean-claude-cli
description	Expert guide for Jean Claude CLI - a sophisticated AI orchestration framework featuring two-agent workflows, event sourcing, coordinator pattern with ntfy.sh, agent note-taking, mailbox communication, and real-time monitoring. Use when user asks about jc commands, workflows, Beads integration, agent coordination, or Jean Claude architecture.

Jean Claude CLI Expert

Expert knowledge for using Jean Claude CLI (jc), a sophisticated AI-driven development workflow orchestration framework.

Architecture Overview

Jean Claude is a two-layer orchestration framework:

1. Agentic Layer: .claude/commands/ slash commands for Claude Code
2. Application Layer: src/jean_claude/ CLI tool with four subsystems:
   • cli/ - Click-based command interface (14+ commands)
   • core/ - Business logic (56+ modules)
   • orchestration/ - Multi-agent workflow engine
   • dashboard/ - FastAPI monitoring UI with SSE streaming

Key Innovations:

• Event-sourced architecture (SQLite + JSON state)
• Two-agent pattern (Opus plans → Sonnet implements)
• Coordinator pattern (agent-to-agent + agent-to-human communication)
• Worktree isolation (parallel execution)

Core Commands

jc init

Initialize Jean Claude in a project (run once).

Creates:

• .jc-project.yaml - Project configuration
• .claude/skills/jean-claude-cli/ - This skill
• specs/ - Workflow specifications directory
• agents/ - Agent working directories

Example:

jc init

jc prompt "description"

Execute a single prompt with Claude Agent SDK.

When to use:

• Quick one-off tasks
• Testing prompts
• Simple code generation without workflow overhead

Options:

• --model opus|sonnet|haiku - Choose model (default: sonnet)
• --stream - Show real-time output
• --raw - Return raw response without formatting

Example:

jc prompt "Add docstrings to all functions" --model sonnet --stream

jc workflow "description"

Run two-agent workflow without Beads tracking.

When to use:

• Ad-hoc feature development
• Complex multi-step tasks requiring planning
• Tasks needing strategic + tactical execution

Phases:

1. Planning (Opus): Analyzes scope, creates feature list in state.json
2. Implementation (Sonnet): Implements features one by one with fresh context
3. Verification: Runs tests, checks completion

Options:

• --initializer-model opus|sonnet - Planning model (default: opus)
• --coder-model opus|sonnet|haiku - Implementation model (default: sonnet)
• --max-iterations N - Max iterations (default: 10)
• --auto-continue - Resume automatically if interrupted

Example:

jc workflow "Add user authentication with JWT tokens" --auto-continue

jc work <task-id>

Execute a Beads task using two-agent workflow.

When to use:

• Working on Beads-tracked issues
• Feature development with external issue tracking
• Tasks requiring team coordination

Process:

1. Fetches task from Beads (bd show <task-id>)
2. Generates spec in specs/beads-{task-id}.md
3. Runs two-agent workflow (Opus plans → Sonnet implements)
4. Updates Beads task status
5. Verifies completion

Example:

jc work beads-abc123

jc note

Agent note-taking system for persistent knowledge capture.

When to use:

• Capturing insights during workflow execution
• Documenting decisions and patterns
• Building institutional knowledge
• Sharing context across workflows

Subcommands:

• jc note add "content" - Add note (agents can use this proactively)
• jc note list - View all notes
• jc note list --workflow <id> - Notes for specific workflow
• jc note search "query" - Search note contents

Agent Usage: Agents executing in workflows can proactively use the note-taking API to:

• Document architectural decisions
• Record trade-offs considered
• Share discoveries with future workflows
• Build project-specific knowledge base

Example:

# Human usage
jc note add "Using SQLAlchemy 2.0 async patterns for all DB access"
jc note list --workflow a3b4c5d6

# Agent usage (within workflow)
await note_taking_api.create_note(
    content="Discovered circular import in auth module",
    workflow_id=workflow_id,
    category="architecture"
)

jc prime

Gather project context efficiently.

When to use:

• Starting work in new codebase
• After long break from project
• Before planning major changes

Process: Uses fast Haiku model to explore codebase and return condensed summary (~500 words) covering:

• Tech stack
• Entry points
• Testing setup
• Architecture patterns

Options:

• --raw - Return raw markdown without formatting
• --output <file> - Save to file

Example:

jc prime --raw

jc status

Check workflow status and progress.

When to use:

• Monitoring running workflows
• Checking feature completion
• Debugging stuck workflows

Options:

• <workflow-id> - Specific workflow (default: latest)
• --json - JSON output for scripting

Example:

jc status                    # Latest workflow
jc status a3b4c5d6          # Specific workflow
jc status --json | jq .      # Programmatic access

jc logs

View workflow event logs.

When to use:

• Debugging workflow issues
• Auditing agent actions
• Real-time monitoring

Options:

• <workflow-id> - Specific workflow (default: latest)
• --follow - Follow mode (real-time, like tail -f)
• --level debug|info|warning|error - Filter by log level

Example:

jc logs --follow              # Real-time monitoring
jc logs a3b4c5d6             # Specific workflow
jc logs --level error         # Errors only

jc dashboard

Launch web monitoring dashboard.

When to use:

• Visual workflow monitoring
• Team collaboration
• Real-time progress tracking

Features:

• Live SSE streaming (no polling)
• Event timeline visualization
• Feature progress tracking
• Cost and duration metrics
• Multi-workflow monitoring

Example:

jc dashboard                  # Launches on http://localhost:8000

Access dashboard in browser to see:

• Real-time event stream
• Workflow state (features, phases, costs)
• Agent communication logs
• Test results and validation

jc migrate

Update existing project to latest Jean Claude version.

When to use:

• After upgrading Jean Claude
• Adopting new features
• Syncing project structure

Process:

• Creates missing directories
• Adds new slash commands
• Updates jean-claude-cli skill with latest features
• Updates CLAUDE.md section
• Does NOT overwrite existing files

Options:

• --dry-run - Preview changes without applying

Example:

jc migrate --dry-run          # Preview
jc migrate                    # Apply updates

Two-Agent Pattern

Jean Claude's core innovation: strategic planning with tactical execution.

Architecture

1. Initializer Agent (Opus):

   • Analyzes scope ONCE
   • Creates feature list as JSON
   • Writes to agents/{workflow-id}/state.json
   • Expensive but thorough

2. Coder Agent (Sonnet):

   • Loops through feature list
   • Implements ONE feature per iteration
   • Gets FRESH context each iteration (prevents bloat)
   • Verifies after each feature
   • Cheap and efficient

3. Shared State:

   • agents/{workflow-id}/state.json - Single source of truth
   • Features: not_started → in_progress → completed
   • Phases: planning → implementing → verifying → complete
   • Event log: agents/{workflow-id}/events.jsonl

Benefits

• Cost-effective: Expensive model only for planning
• Context management: Fresh context per feature prevents bloat
• Quality: Strategic planning + focused execution
• Resumable: Crash recovery from state.json
• Auditable: Complete event log

State Management

WorkflowState structure (agents/{workflow-id}/state.json):

{
  "workflow_id": "a3b4c5d6",
  "features": [
    {
      "id": "feat-1",
      "name": "Add JWT authentication",
      "status": "completed",
      "description": "...",
      "verification": "..."
    }
  ],
  "current_phase": "implementing",
  "costs": {"initializer": 0.45, "coder": 1.23},
  "session_ids": ["sess-1", "sess-2"]
}

Event Sourcing Architecture

Jean Claude uses event sourcing for complete auditability and crash recovery.

Event Store

Dual persistence:

1. SQLite database (.jc/events.db) - Centralized, queryable
2. JSONL files (agents/{workflow-id}/events.jsonl) - Per-workflow append-only log

Event types:

• workflow_started, workflow_completed
• feature_started, feature_completed
• agent_invocation, agent_response
• test_run, validation_check
• error_detected, blocker_detected
• message_sent, message_received
• note_created, note_updated

Snapshots

Every 100 events, system creates snapshot for bounded replay:

• Current workflow state
• Feature completion status
• Agent session history
• Cost accumulation

Benefits:

• Instant crash recovery
• Complete audit trail
• Deterministic replay
• Performance optimization

Querying Events

# Via CLI
jc logs a3b4c5d6 --level error

# Via SQLite
sqlite3 .jc/events.db "SELECT * FROM events WHERE workflow_id='a3b4c5d6'"

# Via dashboard
jc dashboard  # Real-time event stream

Coordinator Pattern

Jean Claude implements hierarchical agent coordination with human-in-the-loop escalation.

Architecture

Main Claude Code (Coordinator)
    ↓
Subagents (Initializer, Coder, etc.)
    ↓
Mailbox Tools (ask_user, notify_user)
    ↓
Coordinator Triage (90% auto-answer, 10% escalate)
    ↓
ntfy.sh (Mobile notifications)
    ↓
Human Response

Mailbox Communication

Location: agents/{workflow-id}/INBOX/ and OUTBOX/

Message format:

{
  "id": "msg-abc123",
  "from": "coder-agent",
  "to": "coordinator",
  "priority": "normal",
  "question": "Should I use SQLite or PostgreSQL?",
  "context": {"feature_id": "feat-2"},
  "created_at": "2026-01-03T12:34:56Z"
}

Priorities:

• LOW - FYI, non-blocking
• NORMAL - Needs answer within 1 hour
• URGENT - Critical decision, blocks progress
• CRITICAL - Safety/security concern

Coordinator Triage

Automatic answers (90%):

• Questions answerable from codebase
• Decisions matching existing patterns
• Simple clarifications

Escalated to human (10%):

• Architectural decisions
• Business logic choices
• Security/compliance questions
• Ambiguous requirements

ntfy.sh Integration

Setup:

# .env configuration
export JEAN_CLAUDE_NTFY_TOPIC="your-escalation-topic"
export JEAN_CLAUDE_NTFY_RESPONSE_TOPIC="your-response-topic"

Notification format:

[project-name] Question from Coder Agent

Workflow: a3b4c5d6
Feature: Add authentication

Question: Should I implement OAuth2 or JWT tokens?
Context: User mentioned "simple auth" but didn't specify protocol.

Reply with: a3b4c5d6: your response

Response format (from phone):

a3b4c5d6: Use JWT tokens for simplicity

Multi-project support: All projects share same ntfy topics. Workflow ID ensures correct routing:

Project A (jean-claude):      a3b4c5d6
Project B (my-api-server):    f8e2a1b9
Project C (website):          2c7d9e4a

# You receive:
[jean-claude] Architecture Question
[my-api-server] Should I add rate limiting?
[website] Use SQLite or Postgres?

# You respond:
a3b4c5d6: Use the pattern from existing code
f8e2a1b9: Yes, add rate limiting
2c7d9e4a: Use Postgres

Polling Pattern

Coordinators poll for responses (not blocking):

max_attempts = 30  # 30 attempts × 10 seconds = 5 minutes
for attempt in range(max_attempts):
    time.sleep(10)  # Poll every 10 seconds

    responses = poll_ntfy_responses()
    matching = [r for r in responses if r['workflow_id'] == workflow_id]

    if matching:
        response = matching[0]['response']
        break
else:
    # Timeout - proceed with default or pause workflow
    handle_timeout()

Why polling?:

• Respects asynchronous nature of mobile communication
• Allows human time to think
• Provides clear timeout behavior
• Shows progress during wait

Beads Integration

Jean Claude integrates with Beads issue tracker for project-wide task management.

Beads Workflow

# 1. Find available work
bd ready                           # Show tasks with no blockers

# 2. Execute task
jc work beads-abc123              # Two-agent workflow

# 3. Close when done
bd close beads-abc123             # Mark complete

Finding Work

bd ready                          # Available tasks (no blockers)
bd list --status=open             # All open tasks
bd list --status=in_progress      # Your active work
bd show beads-abc123              # Detailed task view

Creating Tasks

# Create task
bd create \
  --title="Add feature X" \
  --type=feature \
  --priority=2                    # 0-4 or P0-P4 (0=critical, 4=backlog)

# Add dependencies
bd dep add beads-yyy beads-xxx   # yyy depends on xxx (xxx blocks yyy)

Closing Tasks

bd close beads-abc123                              # Single task
bd close beads-abc beads-def beads-xyz            # Batch (efficient!)
bd close beads-abc --reason="Completed in PR #42" # With reason

Sync and Collaboration

bd sync                           # Push to git remote
bd sync --status                  # Check sync status
bd stats                          # Project statistics
bd doctor                         # Check for issues

Spec Generation

When running jc work <task-id>, Jean Claude:

1. Fetches task details: bd show <task-id>
2. Generates spec: specs/beads-{task-id}.md
3. Renders with Jinja2 template
4. Passes spec to Initializer agent

Spec template (src/jean_claude/templates/beads_spec.md):

# {{ title }}

**Task ID**: {{ task_id }}
**Type**: {{ type }}
**Priority**: {{ priority }}

## Description
{{ description }}

## Acceptance Criteria
{{ acceptance_criteria }}

## Dependencies
{% for dep in dependencies %}
- {{ dep }}
{% endfor %}

Advanced Features

Auto-Continue

Autonomous continuation loop with error recovery.

When to use:

• Long-running workflows
• Unattended execution
• Large migrations/refactorings

How it works:

1. Coder implements feature
2. Runs tests automatically
3. If tests fail:
   • Detects error type
   • Attempts automatic fix
   • Retries (max 3 attempts)
4. If tests pass:
   • Moves to next feature
5. If blocked:
   • Writes message to INBOX
   • Waits for coordinator

Options:

jc workflow "Large refactoring" \
  --auto-continue \
  --max-iterations 50

Error detection:

• Test failures (test_failure_detector.py)
• Ambiguity detection (ambiguity_detector.py)
• Blocker detection (blocker_detector.py)
• Runtime errors (error_detector.py)

Verification System

After each feature:

1. Runs test command (uv run pytest)
2. Parses output for failures
3. If failures:
   • Logs to event store
   • Triggers auto-continue retry
   • Or escalates to coordinator
4. If success:
   • Marks feature complete
   • Proceeds to next feature

Test commands (.jc-project.yaml):

tooling:
  test_command: uv run pytest
  linter_command: uv run ruff check .
  format_command: uv run ruff format .

Worktree Isolation

Future feature: Each workflow executes in isolated git worktree.

Benefits:

• Parallel workflow execution
• No branch conflicts
• Clean commit history
• Easy rollback

Planned structure:

trees/
├── workflow-a3b4c5d6/    # Isolated worktree
│   └── .git              # Linked to main repo
└── workflow-f8e2a1b9/    # Another workflow
    └── .git

Security Hooks

Jean Claude validates all bash commands before execution.

Validation (src/jean_claude/core/security.py):

• Dangerous commands (rm -rf /, dd, mkfs)
• Unquoted variables in rm/mv
• Eval/exec with user input
• Network commands to internal IPs

Override (if needed):

# In workflow state
"security": {
  "allow_dangerous": false,
  "allowed_commands": ["rm -rf node_modules"]
}

Dashboard & Monitoring

Web Dashboard

jc dashboard              # Launches on http://localhost:8000

Features:

• Real-time SSE streaming (no polling, zero latency)
• Event timeline with filtering
• Feature progress visualization
• Cost tracking (per agent, per feature)
• Duration metrics
• Multi-workflow support

Tech stack:

• FastAPI + Uvicorn
• sse-starlette for live updates
• Jinja2 templates
• SQLite event queries

Event Streaming

Server-Sent Events provide zero-latency updates:

// Browser connects to /events stream
const eventSource = new EventSource('/events');

eventSource.onmessage = (event) => {
  const data = JSON.parse(event.data);
  // Update UI in real-time
};

Event types streamed:

• Feature started/completed
• Test results
• Agent messages
• Cost updates
• Error/blocker detection

Cost Tracking

Tracked per workflow:

{
  "costs": {
    "initializer": 0.45,      // Opus planning
    "coder": 1.23,            // Sonnet implementation
    "total": 1.68
  },
  "tokens": {
    "input": 125000,
    "output": 45000
  }
}

Access via:

• jc status --json | jq .costs
• Dashboard cost panel
• Event store queries

Configuration

Project Config (.jc-project.yaml)

directories:
  specs: specs/              # Workflow specifications
  agents: agents/            # Agent working directories (state.json)
  trees: trees/              # Git worktrees (future)
  source: src/
  tests: tests/

tooling:
  test_command: uv run pytest
  linter_command: uv run ruff check .
  format_command: uv run ruff format .

workflows:
  default_model: sonnet      # sonnet|opus|haiku
  auto_commit: true          # Auto-commit after features
  max_iterations: 10

vcs:
  issue_tracker: beads       # beads|github
  platform: github           # github|gitlab

Environment Variables

Authentication:

ANTHROPIC_API_KEY=sk-...           # Optional (not needed for Claude Max)
CLAUDE_CODE_USE_BEDROCK=1          # Use AWS Bedrock backend
AWS_PROFILE=default                # Bedrock credentials
AWS_REGION=us-east-1

Project:

ADW_ISSUE_TRACKER=beads            # beads|github
ADW_MODEL_SET=base                 # base (sonnet) | advanced (opus)
LOG_LEVEL=INFO                     # DEBUG|INFO|WARNING|ERROR|CRITICAL

Notifications:

JEAN_CLAUDE_NTFY_TOPIC=escalation-topic
JEAN_CLAUDE_NTFY_RESPONSE_TOPIC=response-topic

Common Workflows

Feature Development with Beads

# 1. Create task
bd create --title="Add user authentication" --type=feature --priority=2

# 2. Execute with two-agent workflow
jc work beads-auth123 --auto-continue

# 3. Monitor progress
jc logs --follow

# 4. Verify
uv run pytest tests/
jc status

# 5. Close task
bd close beads-auth123

Ad-hoc Feature Development

# Direct workflow without Beads
jc workflow "Add user authentication with JWT" --auto-continue

# Monitor in dashboard
jc dashboard

# Check status
jc status --json

Quick Refactoring

jc prompt "Refactor module X to use async/await" --model sonnet --stream

Large Migration

jc workflow "Migrate from SQLAlchemy 1.4 to 2.0" \
  --initializer-model opus \
  --coder-model sonnet \
  --auto-continue \
  --max-iterations 30

Agent Note-Taking During Workflow

Agents proactively capture knowledge:

# Within workflow execution
await note_taking_api.create_note(
    content="Using Redis pub/sub for real-time features",
    workflow_id=workflow_id,
    category="architecture"
)

Review notes later:

jc note list --workflow a3b4c5d6
jc note search "Redis"

Multi-Project Coordination

Run workflows across projects:

# Terminal 1: Project A
cd ~/projects/jean-claude
jc workflow "Add feature X" --auto-continue

# Terminal 2: Project B
cd ~/projects/my-api
jc workflow "Add feature Y" --auto-continue

# Terminal 3: Project C
cd ~/projects/website
jc workflow "Add feature Z" --auto-continue

Phone receives:

[jean-claude] Architecture Question (workflow: a3b4c5d6)
[my-api] Database choice? (workflow: f8e2a1b9)
[website] Use REST or GraphQL? (workflow: 2c7d9e4a)

Respond with workflow IDs:

a3b4c5d6: Use pattern from existing code
f8e2a1b9: Use PostgreSQL
2c7d9e4a: Use REST for simplicity

Troubleshooting

Workflow Stuck

# Check status
jc status

# View logs
jc logs --follow

# Check for agent questions
ls agents/*/INBOX/

# Read messages
cat agents/a3b4c5d6/INBOX/msg-*.json

# Respond via coordinator or manually

Beads Sync Issues

bd doctor                 # Check for issues
bd sync --status          # Check sync status
bd sync                   # Force sync

Event Store Corruption

# Rebuild from JSONL logs
jc doctor --rebuild-events

# Verify integrity
sqlite3 .jc/events.db "PRAGMA integrity_check"

Dashboard Not Updating

# Check SSE connection
curl http://localhost:8000/events

# Restart dashboard
pkill -f "jc dashboard"
jc dashboard

ntfy.sh Notifications Not Working

# Test topic
curl -d "Test message" ntfy.sh/your-topic

# Check environment variables
echo $JEAN_CLAUDE_NTFY_TOPIC
echo $JEAN_CLAUDE_NTFY_RESPONSE_TOPIC

# Test from CLI
jc test-ntfy "Test escalation"

High Costs

# Check workflow costs
jc status --json | jq .costs

# Use cheaper models
jc workflow "..." --coder-model haiku

# Reduce iterations
jc workflow "..." --max-iterations 5

Key Concepts

• Workflow ID: Unique 8-char UUID per workflow (e.g., a3b4c5d6)
• State JSON: agents/{workflow-id}/state.json - single source of truth
• Feature List: Initializer creates, Coder implements one-by-one
• Context Reset: Fresh context per feature prevents bloat
• Event Sourcing: Immutable event log for auditability
• Coordinator: Main Claude Code instance managing subagents
• Mailbox: INBOX/OUTBOX for agent-to-agent communication
• Escalation: 10% of questions go to human via ntfy.sh
• Snapshot: Every 100 events for bounded replay
• Verification: Tests after each feature
• Auto-continue: Autonomous error recovery loop

Best Practices

1. Use Beads for tracking: All significant work should have task ID
2. Close tasks in batches: bd close task1 task2 task3 (efficient)
3. Enable auto-continue for large tasks: Unattended execution
4. Let coordinator handle questions: Don't interrupt workflows manually
5. Respond promptly to escalations: Coordinators timeout after 30 minutes
6. Monitor via dashboard: Real-time visibility beats polling
7. Review notes regularly: Build institutional knowledge
8. Use appropriate models: Haiku for simple tasks, Opus for complex planning
9. Verify before closing: Tests pass and feature works
10. Keep workflow IDs handy: Needed for multi-project responses

Architecture Deep Dives

Two-Agent State Machine

┌─────────────────┐
│ Workflow Start  │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Initializer     │ (Opus, once)
│ - Analyze scope │
│ - Create list   │
│ - Write state   │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Coder Loop      │ (Sonnet, N times)
│ For each feat:  │
│ - Read state    │
│ - Fresh context │
│ - Implement     │
│ - Test          │
│ - Update state  │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Verification    │
│ - All tests pass│
│ - Features done │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Workflow End    │
└─────────────────┘

Event Sourcing Flow

┌────────────┐     ┌──────────────┐     ┌─────────────┐
│ Workflow   │────▶│ Event        │────▶│ SQLite DB   │
│ Action     │     │ Emission     │     │ + JSONL     │
└────────────┘     └──────────────┘     └─────────────┘
                           │
                           ▼
                   ┌──────────────┐
                   │ Snapshot     │ (every 100 events)
                   │ Creation     │
                   └──────────────┘
                           │
                           ▼
                   ┌──────────────┐
                   │ Dashboard    │ (SSE stream)
                   │ Update       │
                   └──────────────┘

Coordinator Communication Flow

┌─────────────┐
│ Subagent    │
│ needs help  │
└──────┬──────┘
       │
       ▼
┌─────────────────┐
│ ask_user()      │
│ writes to INBOX │
└──────┬──────────┘
       │
       ▼
┌─────────────────┐
│ Coordinator     │────┐
│ reads INBOX     │    │ 90% auto-answer
└──────┬──────────┘    │ from codebase
       │               └─────────────┐
       │ 10% escalate                │
       ▼                             ▼
┌─────────────────┐          ┌──────────────┐
│ ntfy.sh         │          │ write to     │
│ notification    │          │ OUTBOX       │
└──────┬──────────┘          └──────────────┘
       │
       ▼
┌─────────────────┐
│ Human phone     │
│ responds        │
└──────┬──────────┘
       │
       ▼
┌─────────────────┐
│ Coordinator     │
│ polls response  │
│ writes to       │
│ OUTBOX          │
└──────┬──────────┘
       │
       ▼
┌─────────────────┐
│ Subagent reads  │
│ OUTBOX          │
│ continues work  │
└─────────────────┘

Resources

• Documentation: docs/ directory

  • two-agent-workflow.md - Opus→Sonnet pattern details
  • auto-continue-workflow.md - Error recovery loops
  • coordinator-pattern.md - Agent communication
  • event-store-architecture.md - Event sourcing design
  • beads-workflow.md - Issue tracker integration
  • streaming-implementation-summary.md - SSE architecture

• Templates: src/jean_claude/templates/

  • beads_spec.md - Jinja2 template for Beads tasks

• Skills: .claude/skills/

  • jean-claude-cli/ - This skill

• Project Config: .jc-project.yaml - Project-specific settings

---

Remember: Jean Claude orchestrates workflows, not just tasks. The two-agent pattern provides strategic planning with tactical execution. Event sourcing ensures auditability. The coordinator pattern scales agent autonomy while keeping humans in the loop for critical decisions.