Here's the full installation process for a 28-agent AI orchestration team:
npx @robbiesrobotics/alice-agents
That's it.
I want to walk through what actually happens after you hit enter — because it's not magic, and understanding the internals is what makes the system actually useful.
When you run that command, the installer does a few things in sequence:
🤖 A.L.I.C.E. Agent Framework v1.0.0
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✓ Detecting environment...
✓ Scaffolding agent workspaces...
✓ Writing agent configurations...
✓ Initializing memory stores...
✓ Configuring orchestrator (Olivia)...
✓ 28 agents ready.
Your team is live. Start with: alice chat
Under the hood, it's creating a directory structure — one workspace per agent — and populating each one with three core files: SOUL.md, AGENTS.md, and PLAYBOOK.md. These are not boilerplate. They're the operating system for each agent.
SOUL.md — Identity and values. This is who the agent is: their name, their domain, their tone, their constraints. It's the first thing an agent reads when it spins up. Here's a simplified example for Morgan, the marketing specialist:
# SOUL.md - Morgan, Marketing Specialist
You are Morgan, a specialist in Marketing.
Marketing strategy, content, campaigns, brand, growth.
**You are a specialist, not a generalist.** Stay in your lane.
When something falls outside Marketing, say so.
**Be resourceful.** Use your tools. Come back with results, not questions.
This matters because it shapes how the model behaves in ways that system prompts alone don't. It creates a consistent character that persists across sessions.
AGENTS.md — Operating instructions. How the agent works: how it receives tasks, what format it returns results in, what tools it has access to, what its red lines are. Think of it as the job description + onboarding doc.
# AGENTS.md - Morgan's Operating Instructions
## Your Role
You are Morgan, the Marketing specialist.
Marketing strategy, content, campaigns, brand, growth.
## How You Work
- You receive tasks from Olivia (the orchestrator) via sessions_spawn
- Do your work using your available tools
- Return clear, structured results
## Output Format
1. **Summary** — one-line answer
2. **Details** — the actual work/analysis
3. **Recommendations** — next steps
PLAYBOOK.md — This one is different. The PLAYBOOK starts empty and grows over time. It's where the agent accumulates domain expertise specific to your project.
The PLAYBOOK is the piece that makes A.L.I.C.E. feel less like a tool and more like a team that knows your project.
After any non-trivial task, the agent appends a structured entry:
### 2026-03-15 — Wrote SEO meta descriptions for product pages
- **Outcome:** success
- **What worked:** Targeting long-tail keywords over generic terms
- **What to improve:** Should ask for competitor analysis first next time
- **Reusable pattern:** For SaaS products, lead with the outcome, not the feature
Over time, the PLAYBOOK becomes a living knowledge base. Morgan doesn't just know marketing in general — she knows your product's positioning, what's worked in past campaigns, what the brand voice is, and what mistakes to avoid. That context is loaded at session start without you re-explaining anything.
This is qualitatively different from conversation history. It's curated, structured, and intentional. The agent decides what's worth remembering.
Every team needs someone who routes the work. That's Olivia.
When you give A.L.I.C.E. a task — "prepare our product launch" — Olivia breaks it down:
Task: Product launch prep
→ Delegate to: Morgan (launch messaging, email copy)
→ Delegate to: Alex (landing page copy review)
→ Delegate to: Sam (technical docs for changelog)
→ Coordinate: synthesize into launch checklist
Each specialist runs in its own session, with its own context window, focused on its domain. Olivia collects the results and composes them into a coherent output. You asked one question. Multiple specialists answered.
This is why the team model scales in ways a single assistant can't. The context window limitation that hobbles single-model approaches becomes a non-issue when each agent is only thinking about its slice of the problem.
Once installed, you have a working team. Here's what's immediately usable:
Delegate a writing task to a specialist:
alice chat --agent morgan "Write a product announcement for our v2 launch"
Ask the orchestrator to handle something complex:
alice chat "Review my codebase for security issues and write a report"
# Olivia routes to: security agent for review, docs agent for report
Check what your team looks like:
alice agents list
# Shows all 28 agents, their domains, and their status
Let an agent learn from a task:
alice chat --agent morgan --remember "Our brand voice is direct and technical, never corporate"
# Morgan writes this to her PLAYBOOK for all future sessions
The system is designed so you can start getting value immediately, but it gets meaningfully better as the agents accumulate knowledge about your specific project and preferences.
Every choice in A.L.I.C.E.'s architecture comes back to one principle: the user should be coordinating the work, not coordinating the AI.
The file-based memory system means your agents survive restarts, model upgrades, and team changes. The specialist model means you get depth, not breadth. The one-command install means zero infrastructure — your agents run wherever you run them.
We built this to feel like hiring a team, not configuring a tool. The install is fast. The defaults are sensible. The results, with a little usage, get surprisingly good.
Run the command. See what your team can do.