Eight agent tools, you may keep one: the detailed comparison

The rule: one tool manages, the rest serve

Most people I talk to do not suffer from a shortage of tools. They suffer from too many. Eight subscriptions, eight windows, and in the end they are the glue between all of them. That is not a stack, it is a pile.

A stack starts with one decision: one tool takes the task from me, breaks it down, distributes work, and returns proof. Every other tool receives closed tasks from it, or serves it with information. Once there is one manager, the question "which tool" becomes easy: every tool has one job, and if it does not do that job better than the manager, it leaves the stack.

The tournament in the Reel is exactly that test. In every pair I asked: if I may keep only one, which of the two can be the manager? That is why the winners are not necessarily the smartest tools, but the ones that can run alone, hold context, and return proof.

All eight tools in one table

As of September 2026, from each tool's official documentation. Prices in USD per month, and they change.

Claude Code: the agent that works while you are away

Claude Code reads the repo, edits files, runs commands, and connects to your tools. It started in the terminal, and today the same session moves between the terminal, the desktop app, the IDE, the browser and the phone. That is the part most comparisons miss: not where it can run, but that you can start something on the computer, leave, and steer it from your phone.

The features that make it my manager:

• •Skills: a SKILL.md file in a folder, loaded only when called. This is how I turn a process I solved once into something the agent repeats exactly.
• •Hooks: commands that run at fixed events, like before every command or when the agent stops. The difference between an instruction the model may remember and a rule that is enforced.
• •Subagents: each with its own context and tools. I fan out search and evidence gathering, and keep the decision in the main session.
• •Routines and Remote Control: scheduled tasks that run in the cloud on a schedule or a GitHub event, and a local session I steer from my phone with notifications.
• •Layered memory: global, project and local CLAUDE.md, path-scoped .claude/rules/, and auto memory. I have a separate guide on it.
• •MCP and plugins: connections to external tools, and bundles that ship skills, agents, hooks and MCP servers in one install.

What to know before paying: the subscription comes with rolling five-hour and weekly usage windows, shared with chat and Cowork. A long session resends the whole conversation on every request, so short focused sessions are cheaper than one giant one. And CLAUDE.md is context, not law: keep it under 200 lines, and move anything that must be enforced into a hook.

Codex: the executor with the sandbox

In July 2026 the Codex app merged into the new ChatGPT desktop app, which has three experiences: Chat, Work and Codex. The agent itself also runs in the CLI, the IDE extension, the cloud via GitHub and Slack, and from the phone as remote control of an awake computer.

What makes it a workhorse is the sandbox. Three modes: read only, write inside the workspace, full access. Network is off by default with an allowlist. And there is Auto-review: when the agent wants to cross a boundary, instead of asking you the request goes to a reviewer agent. That is exactly what you want from an executor: keep working until the task is done, inside a fence.

• •Parallel cloud tasks: each in an isolated environment with the repo, ending in a PR. codex cloud hands local work to the cloud.
• •PR review: a @codex review comment or automatic review on every PR, with rules inside AGENTS.md.
• •AGENTS.md: an instruction file read from the repo root down to the current folder. /init creates one. Untrusted projects ignore it.
• •Skills, plugins and MCP: skills via $mention, and since CLI 0.146 you can also install Claude Code marketplaces.
• •Headless CI: codex exec with an API key for pipeline runs.

Price and traps: Codex is included in every ChatGPT plan, including Free, with a five-hour window shared with Work. Image generation burns the quota three to five times faster. API-key users get the CLI and IDE, but not the cloud and not GitHub review.

Cursor: an editor with an agent inside

Cursor is an excellent code editor, and the agent inside it has come a long way: file edits, terminal, browser, checkpoints, and Plan, Ask and Debug modes. There are cloud agents that run in parallel on PRs from Slack, GitHub and Linear, and since August 2026 its own code hosting and custom modes.

• •Rules: .cursor/rules/*.mdc files in always, intelligent, glob or manual mode, plus nested AGENTS.md.
• •Multi-model: OpenAI, Anthropic, Google and Cursor's own models, with an Auto mode that routes by cost or intelligence.
• •MCP: .cursor/mcp.json, one-click install from a marketplace.

The traps: Auto mode bills at the routed model's price plus a surcharge per million tokens on third-party models. Cloud agents need a paid plan, write access to the repo and a spend limit, and the cost is opaque until you set one. Usage pool sizes are not published.

Gemini CLI: research with a huge context, and a change you need to know about

Gemini CLI is Google's open-source terminal agent. Two things make it a research machine: the one million token context on Gemini 3 models, and built-in Google Search grounding. When I need to read a huge repo or pull current sources, this is the tool.

• •GEMINI.md: a global, project and per-folder hierarchy with @file imports. You can configure it to read AGENTS.md too.
• •Extensions: bundles of MCP servers, commands, hooks, subagents and skills, installed from GitHub.
• •Checkpoints: a shadow repo in ~/.gemini/history with /restore. Off by default, worth turning on.

The big change: on June 18, 2026 Google stopped serving requests from free individual users and AI Pro and Ultra subscribers in Gemini CLI, and pointed them at the new Antigravity CLI. Who stayed: Code Assist licenses for companies and paid API keys. If you are a solo builder with a regular Google account, check what works for you today first, because some documentation pages still show the old quotas.

Claude Cowork: the agent for people who do not write code

Cowork brings Claude Code's capabilities to non-code work, with no terminal. It works on folders on your computer, on documents, and on connectors like Google Drive, Slack and Microsoft 365. Most sessions run in the cloud. Local folders, local MCP servers, computer use and the built-in browser need the desktop app open.

• •A folder with instructions: read and write in a folder you chose, with its own instruction file and manual, auto or skip approval per connector.
• •Scheduled tasks: run even when the computer sleeps or the app is closed, hourly, daily or weekly. They cannot be tied to a local folder.
• •Built-in browser: since August 2026. Navigates, reads and fills forms. Banking, email and SSO sites are off by default.
• •Sandbox: local code runs in a virtual machine, and connector tokens never enter it.

What is missing: sessions cannot be shared, so a process a team needs to audit does not live here. Computer use is still a preview, slower than connectors, and sometimes needs a second try. And there is no API path, only a subscription. I have a separate guide for operators.

ChatGPT: what is left after everything merged

Since July 2026 ChatGPT is three experiences: Chat for conversation, Work for an agent that produces finished deliverables, and Codex for code in the desktop app. The old agent mode was retired, and Work replaced it.

• •Work: gathers context from plugins and files and produces sheets, slides, documents and sites. Runs for hours. On desktop it can touch local files and apps.
• •Cloud browser: its own remote browser, a sign-in form the model cannot see, and confirmation before a payment or booking.
• •Scheduled tasks with triggers: Gmail, Slack, GitHub PRs, shareable as links.
• •Apps on MCP: one plugin directory, with permission levels from "always ask" to "all actions".

The traps: Codex is not available in the browser or on the phone, only on desktop. The cloud browser is blocked by some sites and does not finish every transaction. A task created inside a project cannot read the project's files. And Work shares the same usage pool as Codex, so a day of reports eats the code quota.

Hermes vs OpenClaw: an agent that learns vs a hub that reaches everywhere

Both are open source, MIT licensed, free, and bring your own API key or local model. Both run on a computer or a VPS and talk to you through WhatsApp, Telegram, Slack and Discord. And still they are opposites.

So which: if you want one personal agent that accumulates experience and runs cron reports on a cheap VPS, Hermes. If you want an always-on assistant on WhatsApp with paired devices, especially if you already pay for Claude Code or Codex and want to reuse that login, OpenClaw. In both cases, a Docker sandbox from day one.

The final: Claude Code vs Codex

Honestly, the choice is not easy, and not because they are similar. Because both run for me every day and do different things. Claude Code is the manager and the architect: it takes the task from me, breaks it down, fans it out to subagents, holds context, and returns proof. Codex is the executor: it takes a closed task, works inside a sandbox, and double-checks what the manager wrote.

If I must keep one, Claude Code. Three reasons that are really one:

• •It is the only one I can leave running alone at night. Routines in the cloud, hooks that enforce boundaries, and Remote Control from the phone when something needs me.
• •It manages the others. Subagents, plugins, MCP, and a call to Codex when I need a different executor or a second opinion.
• •Everything is documented and controlled. Every layer, from CLAUDE.md to a hook, is a text file in the repo I can read, version and restore.

That does not mean Codex is weak. It means an excellent executor is still an executor. The comparison most people make, who writes better code, is not the question. The question is who can be the only one.

Ten tricks for combining the tools

These come out of my day, not a marketing page. Each one solves a problem I hit while running several agents at once.

# in the repo: AGENTS.md is the real file, CLAUDE.md points to it
ln -s AGENTS.md CLAUDE.md
git add AGENTS.md CLAUDE.md

git worktree add ../repo-codex-task origin/main -b codex/task-slug
# Codex works there. Claude Code stays in the main folder.
git worktree remove ../repo-codex-task   # after the merge, never rm -rf

How to choose based on who you are

• •A developer working on a repo every day: Claude Code manages, Codex executes and reviews. Cursor if you must have an editor, and even then the long-running agent runs outside it.
• •A non-technical founder: Cowork with one folder. ChatGPT Work when the deliverable is a document or a deck from the cloud. No terminal, no VPS.
• •An operator who runs processes: Claude Code with routines and hooks, and Cowork for the parts that touch Office files.
• •Someone who wants an assistant on WhatsApp: OpenClaw on a home machine or a VPS, with Docker and skills you approve yourself. Hermes if you want it to learn you.
• •A researcher or anyone working with a huge repo: Gemini with a license or an API key, returning files to the manager.

In every case, the test question is one: if tomorrow only this tool is left, do I still finish the workday? If the answer is no, it is not your manager.

Template: your stack map

Copy it, fill in the blanks, and put it in your manager's instruction folder. If you want the longer version with the selection questions, there is the stack map guide.

# My stack map (September 2026)

## Who manages
One tool that takes the task from me, breaks it down, and returns proof:
- ______ (mine: Claude Code)

## Who executes
Tools that receive a closed task from the manager, in their own worktree:
- ______ (mine: Codex)

## Who researches
A tool that reads a lot and returns a file, not a chat:
- ______ (mine: Gemini with search, or a Claude Code subagent)

## Who reviews
A tool that sees the PR before I do:
- ______ (mine: Codex review on a PR Claude Code opened)

## Who runs while I sleep
- Repo automations: ______ (mine: Claude Code routines)
- Reports to a messaging channel: ______ (mine: Hermes cron)

## Who is not in the stack, and why
- ______

## Three signs the stack works
1. I open the day with a list of proofs, not a list of chats.
2. No tool writes to the same files as another tool at the same time.
3. If a tool disappears tomorrow, one map tells me what replaces it.