The Coder Agents Map — who plays where?
In 2021, GitHub Copilot landed inside the IDE and felt like magic: it completed the line you were typing. A few years later the picture flipped in a way that looks absurd on the surface: the newest, most powerful AI coding tools are not in your fancy IDE — they went back to a 50-year-old interface: the terminal.
Why the terminal?
Simpler than you'd think — and it's the same reason agents succeeded at coding in the first place: the terminal is where everything is text. Reading files? Text. Running tests? Text. Git? Text. Build errors? Text. And what do LLMs do? Text in, text out.
The terminal isn't a regression — it's the LLM's native habitat. Inside an IDE, an agent needs bespoke plugins and APIs for every move. In the terminal, the agent uses exactly the same tools you do, with no middleman.
The map
- Claude Code (Anthropic) — ignited the wave on release in February 2025. Philosophy: one strong agent, explicit permissions, and a harness built for long, focused autonomous runs. Full ecosystem: slash commands, hooks, subagents, MCP, skills.
- GitHub Copilot CLI (GitHub/Microsoft) — Copilot stepping out of the IDE. Its ace: deep GitHub integration (repos, issues, PRs) and bundling with the Copilot subscription millions already have.
- opencode — the open-source player: the same terminal-agent experience with full model freedom — Anthropic, OpenAI, or local models. For those who refuse vendor lock-in.
- Kimi CLI (Moonshot AI) — from China, built around the Kimi models that shook benchmarks at a fraction of the price. Clear pitch: near-frontier capability, different economics.
- Gemini CLI (Google) — Google's entry, leading with a generous free tier and ecosystem integration.
Above all of these sits another layer: orchestrators — systems that run multiple agents together, like Hermes and its kind, distributing work to specialized agents and gathering results. A story of its own.
The twist
"They all look alike — terminal, loop, tools — how do I choose?" Here's the chapter's twist: the real difference isn't features — it's harness and trust. Features equalize within months. What doesn't equalize: the quality of small decisions inside the loop. When does the agent ask before acting? How does it manage context near the limit? What does it do after a mistake? Those show after weeks of use, not in a demo.
Takeaways:
- The terminal won because it's the models' native habitat — text everywhere.
- Choose by the ecosystem you already live in — GitHub-everything → Copilot CLI; model freedom → opencode; strongest integrated harness → Claude Code.
- Theory means nothing without practice — so we built a cheat sheet for every tool, with an interactive terminal that replays the commands live before you install anything.
Open question: with agents evolving this fast, how long does an article like this stay correct? Honest answer: not long. Which is why this site itself runs agents that update its content — the journal about agents is written by agents. Sit with that sentence, then go check the news page.