Codex, Pi, Hermes — Three Species of AI Agents, Not Three Rivals

A question keeps coming up in our engineering group and among the small businesses just starting to adopt AI: "Codex, Pi, or Hermes — which one should I pick?" It sounds simple, but the question is wrong at its root, because these three names do not sit in the same row. They are not three versions of the same thing. They are three different species of AI agent.

Understanding this changes how you choose AI tools for your business — and saves you a lot of money and trial and error. This article breaks down the three species, and more importantly, helps you know which species is for which job.

Asking "which agent is best" is like asking whether a truck or a motorbike is better

There is a useful comparison the engineering community uses for this: asking "is Codex, Pi, or Hermes better?" is like asking "is a truck or a motorbike better?" — the answer is: it depends. Need to haul cargo? The truck wins. Need to weave through narrow streets? The motorbike wins. No species is "best" in absolute terms; there is only a species that fits the right job better.

These three names represent three completely different design philosophies, born to solve three different problems. Let's look at each species in turn.

Codex — the "commercial coding agent" species, results now

Codex is OpenAI's commercial coding agent. This species was born for a single goal: give you coding results immediately, backed by the strongest models available, inside a safe environment.

Codex's strengths are clear: a sandbox that is on by default (the agent runs in an isolated environment and can't touch your system on its own), IDE and cloud integration, and access to the strongest models currently available. If you need an agent that "delivers fast" — tell it to write a piece of code, fix a bug, and get the result back — Codex is the right choice.

Its weakness sits in the very nature of being commercial: vendor lock-in. Closed models, you're tied to the vendor's ecosystem, and you don't control the infrastructure underneath.

Pi — the "minimal harness" species, fast and fully yours

Pi (pi.dev) is a completely different species. It is a minimal agent harness, open source (MIT, roughly 90,000 GitHub stars), written in TypeScript by Mario Zechner — the creator of libGDX. Pi's philosophy is "primitives, not features": it ships only 4 default tools (read, write, edit, bash), and everything else you extend yourself through TypeScript extensions.

Its strength lies in extreme minimalism: an agent loop of only about 418 lines of code, and a system prompt and tools under 1,000 tokens. That makes it extremely fast and cheap on tokens, and you — or the agent itself — can read, understand, and modify the whole machine. Pi also has a session tree that enables branching, time-travel, and self-modification.

But Pi comes with trade-offs: it has no cross-session memory (by design — each session starts from zero), no sandbox by default (you must containerize it yourself), and it is terminal-only. If you want an agent that remembers its work across months, Pi is not that choice.

Hermes — the "persistent assistant" species, durable and company-scale

Hermes Agent is a third species: a persistent personal assistant — a batteries-included daemon, not just a coding tool. Its philosophy is to live with you across many sessions, remember everything you've done, and run an entire "agent company" instead of handling a single coding command.

Hermes's strengths lie in long-term memory, on-demand skills, 20+ messaging gateways (lives across Telegram, Discord, Slack, WhatsApp), kanban orchestration that lets many agents coordinate like an organization, and multi-environment execution. This is exactly why G-Company OS by 5ac.vn is built on Hermes — it is the operational backbone.

Its weakness: Hermes is a monolithic daemon with a broad toolset, so it has a larger attack surface and requires some technical know-how to run optimally.

Three species, one comparison table

To make it easy to picture, here is how the three species stand side by side:

These three species don't compete directly — they complement each other. This is why StandardCompute, a respected AI comparison site, concluded: "different tools for different jobs", and the most common answer in the community is to run both (Hermes for orchestration, Pi for deep coding).

The lesson for small businesses: pick the right species for the right job

So what should your small business do with these three species?

One: don't get swept up in the "which is best" race. Before choosing, say clearly what you need to do. Need a coding agent that delivers results now? Codex. Need an extremely fast terminal tool you own and fully understand? Pi. Need a durable AI assistant that remembers things, runs a whole team of agents, and sends reports over Telegram every morning? Hermes.

Two: think about the harness, not just the model. A key lesson from the research on Pi: the harness affects the cost per task by more than 2x, while quality stays the same. The same model, a different harness, a different cost. So don't just compare "which AI is smarter" — compare the operating harness behind it too. We covered this in depth in our piece on open-source multi-agent harnesses.

Three: ask about ownership and data safety. If your business values control over its data and doesn't want to be locked into one vendor, open source is the right choice — as our article on open-source vs proprietary AI infrastructure explains. Both Pi and Hermes are MIT, putting ownership back in your hands.

Four: don't be afraid to run several species at once. This is exactly what many organizations, including us, do. Hermes handles orchestration — planning, assigning the right agent, remembering everything, sending reports. Pi — or a similar coding harness — handles deep, interactive coding tasks on the terminal. Each species does what it does best.

So which species does 5ac choose?

The short answer: we bet on Hermes as the operational backbone of G-Company OS, and we keep the door open for Pi as a coding harness. This isn't "pick one, drop the other" — it's recognizing that each species serves a different layer of business operations.

If you want to go deeper on how we build this open-source AI infrastructure, our article on open-source AI infrastructure with Hermes is a good starting point.

Conclusion: don't ask "which is best", ask "which species fits the job"

Codex, Pi and Hermes are not rivals — they are three different species of AI agent, each evolved to be great at one thing. The biggest lesson for a small business is not picking the most powerful agent, but knowing clearly what you need, then picking the right species for the right job.

You don't have to buy them all. You just need to pick the species that best matches your problem — and if your problem has many layers, don't hesitate to let several species live and work side by side.

What about you? Does your business need a fast coding agent, a harness you fully own, or a durable assistant that runs operations? Share in the comments — I read them all and will share more practical lessons for small business owners in upcoming posts.