Blog
Mar 27, 2026

Unlocking Multi-Agent "Power Up Mode" with Paperclip

It took 10 days to find the rhythm where AI agents actually produce while you walk away. Here's what worked and what didn't.

Unlocking Multi-Agent "Power Up Mode" with Paperclip

It took about 10 days to unlock what I call Power Up Mode — that moment when you feel like you finally have a rhythm going and your agents are being productive even when you walk away from the computer.

Getting here wasn’t straightforward.

Two False Starts

I tried OpenClaw twice. Two false starts. Then I found Paperclip. This was my jam.

Before Paperclip, I’d been coding in Windsurf, bumping up against credit limits. That was always my limiting factor. The thought of having multiple agents running all day and night sounded very expensive. If I was already maxing out what one agent could do, multi-agent orchestration seemed fantastical.

The Cheap Agent Trap

My first instinct was to go budget: Paperclip with OpenCode, OpenRouter, and cheaper models like MiniMax 2.5.

OpenCode is okay, but it doesn’t provide the context that more mature tools provide. You spend a lot more time hand-holding the agent to get meaningful work done.

As for the cheaper AI models — they just aren’t as fast and reliable as Opus 4.5 or 4.6, or Sonnet 4.6.

Opus 4.5 set the standard for me. If a model can’t be as good as Opus 4.5, I don’t want to waste my time with it.

Here’s the realization that changed everything: less capable, cheaper agents eat your time and probably cost about the same as more expensive, more capable agents. You spend 2x as long doing something. That’s 2x the tokens. You’re paying for cheap agents with your time… and maybe your wallet too.

Going All In on Claude

I scrapped OpenCode and OpenRouter. I reasoned that if Paperclip is going to work effectively, I need capable agents (Opus 4.5 or better) and I need to max them out.

I got a Claude Max subscription. I left Windsurf and went full Claude World — Claude, Claude Desktop, Claude Code.

My agent roster:

  • CEO: Opus 4.6
  • Engineer: Sonnet 4.6
  • Website Designer: Sonnet 4.6
  • Chief Organizer: Sonnet 4.6

Claude Code is installed with plugins, skills, and MCPs connected. The agents run through Paperclip’s orchestration layer.

The Obsidian Unlock

After a couple days of Paperclip, something felt missing. You need a place to have docs — lightweight docs — where you can see output and craft things collaboratively with agents. Google Docs was too heavy. Drive wasn’t going to cut it.

Enter Obsidian.

After researching different memory and context tools, Obsidian was the natural next step. It doesn’t solve the context problem entirely, but it gives me a place to work on ideas, plans, posts, and resources that all my agents can access. It’s really an interface to the knowledge that me and my agents need to function.

We’re crafting a knowledge base together that represents our processes, our collective knowledge, our outputs.

Here’s the key insight about Obsidian: it’s all markdown files. Obsidian is a lightweight UI over a folder of .md files. Why does this matter?

  1. Agents read markdown easily. It’s their native language.
  2. I read markdown easily. It’s a great medium for human-agent collaboration.
  3. It’s future-proof. Memory and context systems are evolving rapidly. I want to let others build and develop them to maturity before I commit. Those .md files will be ready to import into any future system.

The Philosophy

Paperclip will come and go. AI models will come and go. But the context — the knowledge captured in agent files and markdown in Obsidian — that persists.

At some point I’ll find a solid memory system and import everything. Until then, this system works and provides an easy offramp into whatever comes next.

The real compound interest in AI isn’t the tools. It’s the knowledge you build with them.

Know Your Tools

Here’s a key learning that took me a while to internalize: Paperclip is not great at everything I use agents for.

When I want to rapidly build code, I go directly to Claude Desktop and open 2 or 3 sessions. Using Paperclip for that was too tedious. I was constantly opening issues, waiting, opening another issue, waiting, going back to the first issue, waiting. Too much time spent on agents loading context — too much overhead for quick back-and-forth.

Paperclip is better when you can get your agents set up well and give them work that you don’t need done immediately.

Async tasks → use Paperclip. Sync/immediate tasks → use Claude Desktop.

Understanding which tool fits which workflow — that’s the real Power Up Mode.