[Hackathon] feat: Custom Agent Library + Workflow Time Machine

[EmilySun621]

TL;DR: Users create their own specialized AI agents (like a "GPTs Store" for data science workflows) and get Git-like visual version control to track and revert every change — human or AI.

⚠️ Still under testing

---

😤 The Problem

A biomedical researcher asks Texera's generic AI agent to build a workflow. It uses the wrong methodology, skips evaluation metrics, and ignores her lab's standards. She spends an hour fixing it. Her student asks the same question next week and gets different, equally wrong results.

No customization. No consistency. No undo.

---

✨ What We Built

🤖 1. Custom Agent Library

A platform where users create, configure, and share domain-specific AI agents — each with its own expertise baked in.

	Feature	Details
🎛️	Agent creation	Domain, methodology, guardrails, custom rules, knowledge base uploads
🧩	Operator-aware prompts	163 built-in Texera operators injected into the agent's context — agents generate native DAGs, not Python UDFs
🧠	Model selection	Claude Opus for complex tasks, Haiku for simple ones
💬	Per-agent chat history	Each agent maintains its own conversation per workflow
👥	Sharing	Create once, share with your entire lab/team

---

🎬 Demo Walkthrough

1. 🧬 Create a "Diabetes Research Agent" with CRISP-DM methodology + guardrails
2. 💬 Ask it to build a classification workflow → proper DAG appears on canvas using built-in operators
3. ⏳ Open Time Machine → see the 🤖 agent snapshot in the timeline
4. 💥 Make a mistake → one-click revert
5. 🔍 Compare two versions → see exactly what changed

---

🏗️ Architecture

┌─────────────────────────────────────────────────┐
│  🖥️  Frontend (Angular)                         │
│  • Agent Library page (browse/create/edit)      │
│  • Agent selector in chat panel                 │
│  • Time Machine panel + timeline UI             │
├─────────────────────────────────────────────────┤
│  ⚙️  Backend                                     │
│  • agent-service (TS) — system prompt builder   │
│    with operator catalog injection              │
│  • workflow_snapshot table + REST endpoints      │
│    (Amber/Scala)                                │
│  • workflow-history-tool — agent tool for        │
│    version control                              │
├─────────────────────────────────────────────────┤
│  🔒 Texera Core Engine (Amber) — UNMODIFIED     │
│  Everything is additive: new modules,           │
│  new endpoints, new components                  │
└─────────────────────────────────────────────────┘

📊 +3,903 / -533 lines across 44 files · 🔒 Zero modifications to Texera's core engine

---

💡 Why This Matters

Most approaches build one agent that does one thing. We built a platform where users create their own agents that do anything — plus a safety net so they never lose work.

For researchers who don't write code, this is the difference between:

❌ "AI that sometimes helps"

✅ "AI that works the way my lab works, every time."