AI Employee Showdown: Codex, Claude Code, OpenCode, and Hermes Compared for Real Work

Interviewing Four AI Candidates

In 2026, discussions about AI employees center on four names: Codex, Claude Code, OpenCode, and Hermes. Each claims to act as an AI employee. A non-technical friend asked the author to explain the differences, so the author designed a uniform test: organize 20 client contacts, write a follow-up email for each, and track who replied.

Hiring Criteria

The author requires three capabilities from an AI employee:

Batch production — can it handle repetitive writing tasks (outreach emails, weekly reports) in one go?

Access to your data — can it read local files like client lists and Excel reports without manual copy-paste?

Memory — does it remember project status, conversation history, and ownership across sessions?

Candidate 1: Codex

Codex has the lowest barrier: it comes with ChatGPT Plus/Pro subscriptions. It now offers a web panel and a desktop app that can open and edit local files. Given the full test task, Codex reads files, plans, and executes the whole batch with minimal supervision — like a worker who runs an entire pipeline. Ideal for "delegate and collect results later" users. Trade-off: it prefers to run unattended; step-by-step intervention feels less natural.

Candidate 2: Claude Code

Both Codex and Claude Code can access local files; the difference is work style. Claude Code acts like an assistant sitting beside you: it ingests the whole project context, discusses each step ("This client had issue X last time — include it?"), and lets you intervene at any point. It excels at tasks requiring cross-referencing many documents. Downsides: a command-line interface that takes 1–2 days to learn, and quota limits that can halt heavy usage for hours.

Candidate 3: OpenCode

OpenCode is not a single agent but a team framework with a mature ecosystem of 10,000+ pre-built skills (scraping, document processing, report generation). A web console lets you install skills with one click. It can chain multiple specialized agents and supports 50+ underlying AI models. However, it has a history of 100+ vulnerabilities; hundreds of skills are still flagged as risky. The author's rule: only install trusted skills and keep them updated.

Candidate 4: Hermes

Like OpenCode, Hermes provides a team of persistent, memory-enabled agents. Its unique trait: it grows itself . Every solved problem becomes a reusable skill, so the system improves over months. OpenCode follows a fixed workflow; Hermes evolves. Hermes is free, model-agnostic, and now offers a one-click installer (no longer engineer-only). Weaknesses: fewer ready-made skills than OpenCode, requires a dedicated host machine, and no official support.

Side Note: Don't Rush to Pick Just One

The author previously built a custom 8-agent system on Claude from scratch — encountering silent failures, missed scheduled tasks, and no alerts. Lesson: use off-the-shelf tools instead of reinventing the wheel. More importantly, the author treats all four tools as interchangeable modules: today's best brain can be swapped tomorrow without rebuilding the surrounding factory. Avoid vendor lock-in.

Frequently Asked Questions

Chinese support? All four understand and reply in Chinese; only a few one-time backend settings are in English.

Which is safest? Any cloud-backed AI sends data to the vendor's servers. Check company policy; use enterprise plans for sensitive data. For highly confidential material, wait for fully local models.

Pricing (approximate TWD):

Codex: included in ChatGPT subscription (~NT$600/month).

Claude Code: subscription ~NT$600/month, but hits quota limits easily.

OpenCode: software free; pay for the AI model (~NT$600/month or less).

Hermes: software free; same model cost, but lower long-term spend due to skill reuse.

Key point: The software may be free, but the AI brain behind it always costs money (NT$600 ≈ CNY 130–140).

Which One to Start With?

Restricted work computer (cannot install software)? Use Codex web version — runs in browser, zero install, bypasses IT restrictions.

Need AI to read local files? Install a tool on your own machine. Claude Code is best for complex, multi-document tasks. Advanced setup: leave a Mac at home as a server, control it via Telegram from your phone.

Want a persistent team with memory? Choose based on priority: OpenCode for largest skill library and polished console (vet skills, update often); Hermes for a long-term assistant that gets smarter the more you use it. Both need a dedicated host machine.

If you also need integration with internal company systems but lack a spare machine, the author has tested other approaches and will cover them separately. The four options are real; the right choice depends on where you stand today. Your only task now: pick the one that fits your current constraints and start using it.