FOUR CLASSES OF AGENTS
The main mistake when choosing is to compare tools from different categories. It's like arguing whether a screwdriver or a drill is better.
Terminal Agents
They live in the console, work with the project as a developer: read files, edit, run tests, commit. Claude Code, OpenCode, Gemini CLI. They excel at refactoring and multi-file tasks because they see the whole project, not just an open tab.
OpenCode gathered over 160k stars in a year since launch — a sign of how well the format was received.
IDE-Integrated
Cursor, GitHub Copilot, Windsurf. They work where you already are, see the cursor and open file. Best for streaming code writing, weaker on tasks like 'go through the whole project and change architecture'.
Universal Computer Agents
OpenClaw — a fundamentally different beast: it has access not only to code but also to browser, messengers, desktop. After a viral launch in early 2026 it gathered 60k stars in 72 hours. Suitable when the task goes beyond the repository.
Cloud Autonomous
Codex-like services: you set a task, the agent goes off to work in its sandbox and returns a pull request. Convenient for parallel isolated tasks, inconvenient when you need dialogue along the way.
COMPARISON BY KEY PARAMETERS
| Tool | Where it lives | Models | Autonomy | Strong point |
|---|---|---|---|---|
| Claude Code | terminal | only Anthropic | high | quality on large refactorings |
| OpenCode | terminal | 75+ providers | high | freedom to choose model, open source |
| Cursor | IDE (VS Code fork) | several | medium | speed in streaming writing |
| OpenClaw | entire system | any, including Ollama | very high | tasks beyond code |
| Codex | cloud | OpenAI | full | parallel isolated tasks |
MODEL BINDING: AN UNDERVALUED CRITERION
A point that backfires after six months of use. Claude Code works only with Anthropic models — if their pricing policy or availability changes, you have to migrate entirely.
OpenCode supports 75+ providers, OpenClaw is model-agnostic and can work via local Ollama. For Russia this is directly significant: a local model is not blocked and does not require a foreign currency card.
# OpenCode на локальной модели — без API-ключей вообще
export OPENCODE_PROVIDER=ollama
export OPENCODE_MODEL=qwen2.5-coder:14b
opencode
Quality will be lower than top cloud models, but for routine — renames, tests, docstrings — it's more than enough.
WHAT TO CHOOSE FOR SCENARIO
| Scenario | Tool | Justification |
|---|---|---|
| Large legacy refactoring | Claude Code / OpenCode | sees project whole, works multi-step |
| Daily feature writing | Cursor | minimal friction, everything at hand |
| Automation outside code | OpenClaw | browser, files, messengers |
| Bundle of small independent tasks | Codex-like | parallel and isolated |
| Work without external APIs | OpenCode + Ollama | fully local |
| Limited budget | OpenCode + cheap model | DeepSeek/Qwen cost pennies |
WHERE AGENTS BREAK DOWN
Honest part, which marketing doesn't write about.
- Large unfamiliar project. The agent cannot keep a million-line monolith in context. It works in parts, and connections between parts are lost.
- Confident nonsense. It may hallucinate an API of a non-existent library and write coherent code around it. Tests are mandatory, review too.
- Invisible consumption. An autonomous agent in a loop burns tokens in batches. Set limits before launch, not after the bill.
- Security. An agent with shell access can execute a destructive command. Sandbox or at least a separate branch.
WORKING STACK
In practice, experienced teams do not choose a single tool, but combine:
- Cursor — everyday code writing when you're in the driver's seat.
- Terminal agent — large tasks: migrations, refactoring, test coverage.
- Local model — routine and private code that should not leave.
Fresh agent frameworks and their forks regularly appear in the catalog REDDYX — there you can see what actually gains stars and what remains a demo.
FAQ
Will agents replace developers?
For now they shift the load: less typing, more review and task setting. The agent handles clear and repeatable tasks well, but stumbles on the non-obvious and business-context dependent.
Can the agent be run fully locally?
Yes: OpenCode or OpenClaw on top of Ollama. No keys, no blocks, no code leaks. Price — quality lower than top cloud models and hardware requirements.
How many tokens does an autonomous agent consume?
For a medium task — tens of thousands, for a large refactor — easily over a million. Always set a budget limit before launching a long session.
What should a beginner pick?
Cursor: low barrier to entry, familiar interface, errors are visible immediately. Move to terminal agents when you hit tasks like «change everywhere».
