Hermes Agent - AI Orchestrator dengan Integrasi Telegram

1. Apa itu AI Orchestration

AI Orchestration adalah koordinasi multiple AI agents untuk menyelesaikan task kompleks yang terlalu besar atau terlalu beragam untuk single agent. Satu orchestrator memecah task, mendistribusikan subtask ke specialist agents, dan menyatukan hasilnya.

Problem vs Solution

2. Hermes Agent Delegation System

Hermes menggunakan delegate_task() sebagai API utama untuk mendelegasikan work ke sub-agent. Setiap delegation menghasilkan session terisolasi dengan context, tools, dan model config sendiri.

Delegation Modes

Orchestrator vs Leaf Roles

Orchestrator (parent agent) bertugas: menerima user request, memecah task, delegate ke children, merge hasil. Tidak melakukan work berat sendiri.

Leaf agent (child) bertugas: eksekusi spesifik (coding, research, scraping). Punya toolset terbatas, context terisolasi, tidak bisa delegate lebih lanjut kecuali max_spawn_depth mengizinkan.

Configuration

# ~/.hermes/config.yaml
delegation:
  max_iterations: 50          # max tool calls per sub-agent session
  max_concurrent_children: 3  # paralel sub-agents
  max_spawn_depth: 2          # orchestrator -> child -> grandchild (max)

# Setiap sub-agent bisa override model:
delegate_task(
  task="Research competitors",
  model="claude-sonnet-4-20250514",
  tools=["web_search", "web_extract"],
  max_iterations=20
)

3. Telegram Integration

Hermes berjalan sebagai Telegram bot melalui gateway layer. User berinteraksi via chat biasa, agent merespons dengan full tool access.

Gateway Setup

# Install gateway
hermes gateway install

# Configure (set bot token, allowed users)
hermes gateway setup

# Start gateway
hermes gateway start

# Check status
hermes gateway status

# View logs
hermes gateway logs

Environment Variables

Bot API vs User API

4. Multi-Agent Patterns

Hermes mendukung beberapa pattern koordinasi multi-agent. Pilih berdasarkan struktur task, bukan preferensi aesthetic.

Fan-Out Pattern

Orchestrator memecah task identik menjadi N subtask paralel. Cocok untuk: batch processing, multi-source research, parallel testing.

Pipeline Pattern

Output agent A menjadi input agent B. Cocok untuk: research -> outline -> draft -> review -> publish.

Supervisor Pattern

Satu supervisor mengawasi quality output worker agents, bisa retry atau redirect. Cocok untuk: high-quality output yang butuh review loop.

Hub-Spoke Pattern

Satu central agent (hub) memiliki akses ke specialist agents (spokes). Hub routing berdasarkan intent classification.

5. Telegram-Specific Features

Topic Routing (Forum Mode)

Telegram grup bisa diaktifkan sebagai forum dengan topik terpisah. Hermes bisa route pesan ke agent berbeda berdasarkan message_thread_id.

Example Topic Routing Table

Cron Jobs

Hermes mendukung scheduled tasks yang berjalan tanpa user trigger. Bisa full agent mode atau script-only (tanpa LLM).

# Server monitoring setiap 5 menit (script-only, hemat token)
*/5 * * * * hermes cron run server-health --no_agent

# Daily digest jam 9 pagi (full agent mode)
0 9 * * * hermes cron run daily-digest

# Weekly report setiap Senin
0 10 * * 1 hermes cron run weekly-report

Script-only mode (no_agent=True)

Untuk task mekanis yang tidak butuh reasoning (health check, disk usage, restart service), script-only mode menjalankan bash/python script tanpa LLM. Token cost = 0.

Skills, Memory, dan Kanban

Skills

Skills adalah procedural memory dalam format markdown. Disimpan di ~/.hermes/skills/. Setiap skill berisi: trigger conditions, numbered steps, pitfalls, verification. Agent auto-load skill yang relevan sebelum execute.

Memory

Session-based dan persistent memory. Agent bisa menyimpan context antar session (memories/ directory). Berguna untuk preference, project context, learned patterns.

Kanban (SQLite-backed)

Task board built-in untuk koordinasi multi-agent. Mendukung 8 coordination patterns:

Shared Context Flow

Ketika bot-ke-bot communication tidak mungkin, gunakan shared context file:

# shared-context.json - flow states:
planning -> build_needed -> building -> test_needed -> testing -> done

# Orchestrator menulis:
{"task": "build login page", "status": "build_needed", "assigned_to": "builder-bot"}

# Builder bot membaca, menulis:
{"task": "build login page", "status": "test_needed", "output": "/tmp/login.html"}

6. Real-World Use Cases

Customer Support Automation

Telegram bot sebagai first-line support. Agent classify intent, answer FAQ dari knowledge base, escalate ke human jika confidence rendah. Skill: firecrawl-knowledge-ingest untuk build knowledge base dari docs.

Content Pipeline

Research agent gather sources, writer agent draft, editor agent review, publisher agent format dan post. Pipeline pattern dengan kanban tracking. Setiap stage checkpoint ke shared context.

Code Review

Telegram topic #code-review menerima PR link. Agent fetch diff, run multi-dimensional review (security, performance, style), post inline comments. Skill: claude-code-review.

Research Assistant

Deep research dengan fan-out: 9 angle research (economic, technical, social, ...) dijalankan paralel, hasil di-synthesize oleh orchestrator. Skills: deep-research, firecrawl-deep-research.

DevOps Monitoring

Cron-driven monitoring: health check setiap 5 menit, alert ke Telegram topic #monitoring jika anomali. Script-only mode untuk checks, agent mode hanya untuk alert formatting dan root cause analysis.

Store Automation

Telegram digital store: catalog management, order processing, payment tracking. Skill: telegram-digital-store. SQLite-backed product DB, inline keyboard untuk browsing.

7. Architecture Patterns

Single Bot Architecture

Satu bot token, satu Hermes instance. Semua routing via topic/thread_id atau intent classification. Simplest setup, cocok untuk personal use atau small team.

# Config
TELEGRAM_BOT_TOKEN=your_token
TELEGRAM_ALLOWED_USERS=user_id_1,user_id_2

# Routing: topic-based
# All tools available to single bot

Multi-Bot Shared Context

Multiple bot tokens, setiap bot punya Hermes instance terpisah. Komunikasi via shared context file. Cocok untuk: specialization (coding bot vs research bot) dengan different model configs.

# Bot 1 (orchestrator): delegates, routes, merges
# Bot 2 (builder): coding, file operations
# Bot 3 (researcher): web search, scraping

# Shared context di /tmp/hermes-shared/
# File watcher trigger coordination

Cron-Driven Architecture

Tidak ada persistent bot. Cron jobs trigger Hermes agent secara periodik. Cocok untuk: monitoring, reporting, batch processing yang tidak butuh real-time response.

# Script-only (no LLM cost):
*/5 * * * *  hermes cron run disk-check --no_agent
0 */6 * * *  hermes cron run backup --no_agent

# Agent-mode (LLM reasoning):
0 9 * * *    hermes cron run daily-report
0 8 * * 1    hermes cron run weekly-summary

Event-Driven Architecture

Webhook-triggered: external events (GitHub push, form submission, payment callback) trigger Hermes task. Cocok untuk: CI/CD integration, automated workflows. Combine dengan background delegation.

8. Comparison with Alternatives

9. Best Practices

Delegation Rules

Multi-Agent Rules

Anti-Patterns

10. Limitations

Token Costs

Setiap delegation = new session = full system prompt + context. Orchestrator dengan 5 children bisa menghabiskan 5-10x token dibanding single agent. Mitigate: constrain max_iterations, use execute_code untuk task mekanis, use script-only cron jobs.

Rate Limits

LLM providers: RPM/TPM limits berbeda per provider. Fan-out 10 concurrent agents bisa trigger rate limit. Mitigate: max_concurrent_children=3, stagger batch tasks.

Telegram: Bot API ~30 msg/detik per grup. User API FLOOD_WAIT unpredictable. Mitigate: queue messages, respect retry-after header.

Context Window

Model context window (128K-200K) terasa besar tapi cepat habis dengan tool outputs (file contents, search results, code blocks). Sub-agent yang membaca file besar bisa kehabisan context sebelum selesai. Mitigate: chunk reads (offset/limit), summarize before delegate.

Telegram Platform Limits

Security