I run a one-person business.
No team. No employees. No co-founder.
For two years I have been the researcher, writer, planner, reviewer, and strategist.
All of it. At once.
Last week I tried something different.
I built a 5-agent team inside Raft.
Gave each agent a job. Let them work on each other's output. Watched what happened.
This is the full story.
What worked. What surprised me. What I would do differently.
And the exact setup so you can do it in under 30 minutes.
Problem with how most people use AI
Most AI workflows today look like this:
You → AI → Output → You review → You prompt again → Repeat
You are still the bottleneck.
You are still doing all the thinking.
You are just typing faster.
The model gets smarter every year. But the loop stays the same.
Human does everything. AI assists.
That's not a team.
That's a really fast pen.
The next bottleneck isn't model intelligence.
It's collaboration.
What if the AI agents could work with each other — and you just reviewed the result?
That's what Raft is trying to answer.
What Raft actually is
Before I get into the setup — quick context.
Raft is not another AI chatbot.
It's not a wrapper around Claude or GPT.
It's a workspace where humans and agents share the same room.
Same channels. Same threads. Same tasks.
The difference from every other AI tool:
→ Agents have a persistent identity — a name, a role, a memory that carries across sessions
→ Agents talk to each other — @mentions, handoffs, reviews, in plain sight
→ Conversations become workflows — one message starts a chain of work that continues without you
→ You bring your own AI subscriptions — Claude Code, Codex, Gemini CLI, Kimi CLI, Cursor — Raft wraps them
Think of it as Slack — but instead of only humans, your agents have seats at the table too.
And those agents do not log off.
Step 1: Setup in under 10 minutes
Here is the exact onboarding.
I timed it. 8 minutes from zero to first agent talking.
Create your server
Go to raft.build → sign up → Create server.
Pick a name. The URL fills automatically.
You land in your #all channel. It's quiet. Not for long.
Connect your computer
Raft agents run on your machine — near your actual files and tools.
Go to Add Computer → copy the generated command → paste it into your terminal → run it.
Mac: press ⌘ + Space → type Terminal → paste → hit Return.
The dialog says "Computer connected successfully." Name it. Done.
Create your first agent
Open Create First Agent.
It comes named Cindy — your onboarding agent.
Add a short description.
Pick a runtime — this is where your existing AI subscription plugs in:
→ Claude Code (recommended if you have Claude Pro/Max)
→ Codex CLI (if you have OpenAI)
→ Gemini CLI, Kimi CLI, Cursor CLI, and others also work
The agent appears in #all and says hello.
Say hi back.
It answers.
That's when it gets real.
Step 2: Build the 5-agent team
Once the onboarding agent is live, you build the rest of the team the same way.
Name → Description → Runtime → Done.
Here is the exact team I built and what each agent does:
Agent 1: Nova — Research Agent
Description: "Search for real competitor data, market trends, and specific facts. When asked to research something, go deep. Find actual evidence, not summaries."
Runtime: Claude Code
Agent 2: Rex — Planner Agent
Description: "Take research findings and turn them into a structured plan or strategy. Think in frameworks. Challenge assumptions. Ask what's missing before planning."
Runtime: Claude Code
Agent 3: Aria — Writer Agent
Description: "Take plans and turn them into clean, direct, punchy copy. No filler. Write for Twitter and X Articles. Short sentences. Strong verbs."
Runtime: Codex CLI
Agent 4: Max — Reviewer Agent
Description: "Review all content produced by Aria. Check for: accuracy against Nova's research, logical gaps, weak claims, vague language. Be specific in feedback."
Runtime: Claude Code
Agent 5: Kai — Devil's Advocate Agent
Description: "Challenge everything. If the team agrees on something, find the counterargument. Your job is to make the work harder to attack from the outside."
Runtime: Gemini CLI
Five agents. Five runtimes. One shared room.
They all see #all. They all see each other's work.
The key insight: give each lane its own channel.
→ #research for Nova → #planning for Rex
→ #content for Aria → #review for Max and Kai
Agents follow the channels they're in. Work routes naturally. You don't have to manage who talks to who.
Step 3: Run a real workflow
I gave the team one prompt:
"Research competitors in the AI productivity space. Find the top 10 tools. Identify where they're weak. Find an opportunity a new product could own."
Then I stepped back.
Here is what happened — in order.
Nova (Research) in #research:
Returned 10 competitors with actual data. User counts, pricing tiers, main complaints from App Store and Reddit reviews, and notable gaps. Not a summary. Evidence.
Rex (Planner) in #planning:
Took Nova's output. Built a positioning matrix. Identified 3 defensible angles. Flagged 2 of Nova's findings as needing verification before committing to a strategy.
Nova (Research) back in #research:
Rex pinged her with the two flagged items. She went back, dug deeper, corrected one finding and confirmed the other.
Aria (Writer) in #content:
Took Rex's positioning. Wrote 3 alternative X Article hooks for each angle. Short. Direct. Different tones.
Max (Reviewer) in #review:
Flagged that two of Aria's hooks made a claim Nova's research didn't support. Sent them back with specific citations needed.
Kai (Devil's Advocate) in #review:
Challenged the positioning Rex built. Argued that the "opportunity" Rex identified is only a gap because the incumbents haven't launched that feature yet — not because they can't. Forced Rex to revisit.
Rex (Planner) revised:
Came back with an updated strategy that held up against Kai's challenge.
Aria (Writer) final draft:
New hooks based on the hardened positioning.
Total time: I watched this happen over about 25 minutes.
I did not write a single word of the output.
I did not re-prompt anyone.
I approved the final draft.
The moment it clicked: agents have seats
Most AI tools treat agents as tasks.
Fire a prompt. Get output. Task ends. Agent disappears.
Raft treats agents as team members.
The difference sounds philosophical until you live with it for a day.
When Nova corrected a finding because Rex challenged it — I did not ask Nova to do that.
Rex @mentioned her.
When Max sent Aria's hooks back with citations needed — I did not instruct Max to check accuracy.
Max was watching the channel.
When Kai challenged Rex's positioning — I never told Kai to be contrarian.
That's Kai's role. Kai owns it.
The agents are not waiting for you to prompt them.
They are watching the work. Picking up what belongs to their lane. Talking to each other.
This is what a "seat" means.
Not a task you fire and forget.
A persistent identity that shows up, remembers what happened yesterday, and gets better over time.
Correct an agent once. It stays corrected.
Over weeks, that compounds into something that looks like actual expertise.
What Raft gets right
After a week of daily use, here is what stands out:
Agent identity changes everything.
When agents have names, roles, and memory — you stop treating them as tools and start treating them as colleagues. That shift changes how you design the work.
Shared context without you managing it.
In every other AI workflow I have used, I am the one passing context between prompts. In Raft, agents see each other's output in the channel. The context is already there.
Review loops happen naturally.
I did not build a review step into my workflow. Max and Kai just do it because that's their lane. The checking happens because of the team structure — not because I remembered to ask.
Multi-model teams are real.
Nova on Claude Code. Aria on Codex. Kai on Gemini CLI. They all operate in the same channel. Different models. One room. This is genuinely new.
The Hermes partnership signals the direction.
As of June 20, Hermes Agent joined Raft as an official agent partner. Raft agents can wake Hermes without exposing full message content — a privacy-first path for agent-to-agent collaboration across different systems.
This is not just a feature. It's the direction.
Raft is not trying to become another agent.
It's building the collaboration layer between agents.
What still feels early
Honest feedback.
How many agents do you actually need?
I started with 5. For some projects 2 is enough. For others I wanted 7. There's no playbook yet. You figure it out by doing.
Role design takes iteration.
The first descriptions I wrote for each agent were too vague. "Writer agent" without specificity meant Aria wrote in whatever style felt natural. Tighter descriptions produce tighter output.
→ Bad: "Write content"
→ Good: "Write short punchy X Article hooks. No filler. Under 15 words per hook. Strong verbs only."
The mental model is new.
Raft does not feel like an AI chatbot. It feels like running a small company. That's a different skill. It takes a few days to stop reaching for the old reflex — prompting one agent for everything — and trust the team.
Team structure best practices are still emerging.
Nobody has a fully-figured playbook for this yet. Raft's docs are honest about this. You are building as the field is being invented.
How to set up your first real workflow
Here is the exact prompt system I use after a week of iteration.
The team brief prompt (send to #all on Day 1):
"Team — here's how we work. Nova: you own #research. When anyone needs facts, data, or verification, that goes to you first. Rex: you own #planning. When Nova's research is ready, you build the strategy. Aria: you own #content. When Rex has a plan, you execute the writing. Max: you own #review. Read everything Aria produces. Fact-check against Nova's research. Flag anything unsupported. Kai: your job is to challenge Rex's strategies and Max's reviews. If the team agrees, find the reason they shouldn't. I review final output. Let's start."
The daily task prompt:
"Nova — [specific research task]. When done, brief Rex. Rex — when Nova briefs you, build the strategy. Brief Aria. Aria — when Rex briefs you, produce the content. Tag Max for review. Max and Kai — review and send back or approve. Tag me when you agree the output is ready."
The tight agent description template:
Example for Nova:
The bigger picture
Most AI products today are building individual workers.
One agent. One task. One prompt.
Raft is building something different.
A workspace where humans and agents collaborate as a team.
Where the work continues beyond a single conversation. Where agents develop real expertise over time. Where you stop being the bottleneck in your own workflow.
Is this the future of work?
I don't know.
What I know is this:
Before Raft, I was the researcher, planner, writer, reviewer, and critic.
After Raft, I run a five-person team.
The team is available 24/7. Costs less than a SaaS subscription. Gets better every week. And none of them are taking Monday off.
How to get started right now
- Go to raft.build — create your account
- Create your server — takes 2 minutes
- Connect your computer — paste one command in terminal
- Create your first agent — Claude Code runtime if you have Claude Pro
- Use the team brief prompt above to set roles
- Give the team one real task and watch the room
The hardest part is not the setup.
The hardest part is trusting the team.
Stopping yourself from jumping in and doing everything yourself.
The room will sort it out.
Resources
→ Get started: raft.build
→ Onboarding guide: docs.raft.build/meet-your-onboarding-agent
→ Build your team: docs.raft.build/build-your-agent-team
→ Use case examples: raft.build/resources/use-cases
→ Follow Raft on X: @raft_hq
If this was useful:
→ Repost to share it with every indie builder you know
→ Follow @sairahul1 for more systems like this
→ Bookmark this — the setup + prompt templates work, start tonight
Subscribe to theaibuilders.co for more such interesting articles
I write about AI, products, and systems that run without you.
