Everyone should have ten thousand agents working for them.
Not one assistant you take turns with — a fleet. Agents that keep working while you sleep, on machines you already own, carrying everything you have taught them, and reporting back when they are done. That is what PandaNpc is being built for.
Why it is hard today
Most people top out at one or two agents at a time. Not because the models are too weak — because the overhead around each agent is too high. Every new agent has to be started somewhere, taught your conventions again, watched while it works, and collected from by hand. Multiply that by ten and you are the bottleneck. Multiply it by ten thousand and the idea collapses.
So the work is not "a better chat box." It is removing the per-agent overhead, one piece at a time:
- Somewhere to run them — cheap, private compute you already have, not a rented sandbox.
- One memory they all share — teach it once, every agent knows it, on every machine.
- Supervision that does not scale linearly — see many at a glance, step in only when asked.
- Hands beyond the terminal — an agent that can only edit files can only do a fraction of the work.
What already works
Somewhere to run them
- PandaPaw — one command turns any machine you own into an agent host. It installs the engines, joins the machine to your account, and keeps them running. No ports to open, no tokens to paste.
- Three engines, same setup — Claude Code, Codex and PandaCode all run through the same bridge, so adding a different agent is not a different project.
- Your machines form a private network — laptops, servers, home desktops and VMs reach each other directly, so an agent can run wherever the work actually is.
- Bring your own backend — point an engine at a cheaper model provider when the task does not need the expensive one. Running many agents only makes sense if each one is cheap.
One memory they all share
- Agent Center — your rules, skills, memories and agent profiles live with your account and sync to every machine automatically. Write a lesson down once and every agent on every machine has it, in the format its own engine expects.
- Notes — a knowledge base your agents read from and write to, so what one agent learns does not die with its session.
- MCP server — agents reach your notes, models and scheduled tasks through a documented interface instead of scraping screens.
Supervision that does not scale linearly
- Watch from anywhere — iOS app, Windows / macOS / Linux desktop, any browser, or a Chrome side panel. Live output, tool approvals and code changes stay in sync across all of them.
- Step in only when needed — an agent runs unattended and pauses for you when it hits something that needs a decision.
- Hand work to other people — share a live connection so a teammate can drive one of your agents, revocable and time-limited; or share a read-only transcript when they just need to see what happened.
- 29 languages — the whole product, not just the marketing pages.
Hands beyond the terminal
- Browser control — agents drive a real browser with real clicks, including pages that reject synthetic input.
- Remote desktop control — agents see and operate the screen of another machine you own.
- Scheduled work — tasks fire on a schedule and run against a machine of your choosing, with no one at the keyboard.
- Claw — a workspace for the conversational half of the job, alongside the agents doing the mechanical half.
How the pieces fit
Every agent runs on hardware you control. The relay only carries messages; your code, your keys and your files never leave your machines.
Where this is going
Getting from a handful of agents to a fleet is mostly a supervision problem now, not a capability problem. The work ahead is making many concurrent agents legible — knowing at a glance which ones are stuck, which are waiting on you, and which quietly finished — and making it routine for one agent's output to become the next one's input without a human copying it across.
Getting started
New here? The Quick Start Guide gets your first agent running in a few minutes.