FAQ
Frequently Asked Questions
General
Q: What is Agentic Engineering?
A: Agentic Engineering is a software development methodology where AI Agents play a central role in the entire engineering lifecycle - from design to deployment to operations.
Q: Why 1000 Agents?
A: 1000 is a sweet spot that provides:
- Visual impact and clear communication
- Actual productive capacity
- Manageable complexity
- Cost-effective scale
Q: Can I run this locally?
A: Yes! The MVP can run on a single machine with Docker. Full 1000-Agent scale requires a Kubernetes cluster.
Technical
Q: What LLM models are supported?
A: The platform is model-agnostic. Currently optimized for:
- Alibaba Cloud: Qwen3.5-Plus, Qwen3-Max
- OpenAI: GPT-4, GPT-4-Turbo
- Google: Claude, Gemini
Q: How do Agents communicate?
A: Agents communicate through:
- Task queues (for work assignment)
- Shared state (in file system)
- Direct messaging (via Orchestrator)
Q: What happens when an Agent fails?
A: The Cage health monitor detects the failure and:
- Saves current state to persistent storage
- Attempts automatic recovery (restart)
- If recovery fails, escalates to human operator
- Reassigns pending tasks to other Agents
Cost
Q: How much does it cost to run 1000 Agents?
A: Approximately $464,000/month, broken down as:
- Compute: $285,000
- Tokens: $144,000
- Storage: $15,000
- Overhead: $20,000
Q: Can I reduce costs?
A: Yes! Strategies include:
- Use spot instances (60-70% savings on compute)
- Reserved capacity (30-40% savings)
- Token budgeting (20-30% savings)
- Idle detection and scale-down (15-25% savings)
Security
Q: How is data isolated between Agents?
A: Each Cage has:
- Dedicated Kubernetes namespace
- Isolated persistent storage
- Network policies restricting cross-Cage access
- Separate service accounts and credentials
Q: Can Agents access external systems?
A: Yes, but with strict controls:
- Whitelisted external APIs only
- Rate limiting and quotas
- Audit logging of all external calls
- Human approval for sensitive operations
Migration
Q: How long does mono-repo migration take?
A: Based on 10-repo experiment:
- Analysis: ~2 hours per repo (parallel)
- Planning: ~1 day for 400 repos
- Execution: ~2-4 weeks (phased approach)
Q: What’s the risk of migration?
A: Key risks and mitigations:
- Build complexity: Gradual migration, parallel testing
- Agent coordination overhead: Layered scheduling, batch processing
- State consistency: File locks + transaction logs
- Human acceptance: Gradual automation,保留 approval points