Claude has already won. Copilot is dead in the water. Cursor is Elon’s fantasy play.
Every one of those takes is pontification dressed in the robes of an oracle.
This sounds like a crisis. It’s a format war. And we know how format wars end.
The harness conversation feels urgent because the AI tooling space is moving fast and the stakes feel high. Both of those things are true. Neither of them means you need to pick the right answer right now, or that the answer you pick today is permanent. The people telling you otherwise are not oracles. They’re spectators with opinions.
We have been here before. Multiple times. And the pattern is consistent enough that it tells us something useful about where this goes.
Flexibility wins. Every time.
The common thread across every major technology format war is not that the technically superior option won, or that the established player won, or even that the best-funded player won. The winner was the more flexible, more open option. The one that met users where they already were instead of demanding they start over.
MS-DOS and the IBM PC. IBM built the PC with off-the-shelf components and a third-party OS. Any manufacturer who could reverse-engineer the BIOS could clone it legally. Compaq did it in 1982. Proprietary alternatives like the Amiga, Atari ST, and Apple II became niche products or disappeared entirely. Open architecture won on distribution, not on technical merit.
Git. Git could import your entire SVN history on day one. Migrating didn’t mean abandoning your past. That single decision lowered the switching cost to nearly zero and let developers try Git without risk. SVN never recovered.
Google Docs. Google Docs opens .docx files natively. Microsoft Word does not open Google Docs files. The challenger met the incumbent on the incumbent’s turf and made staying with Word feel like the harder choice. That asymmetry still compounds.
Spotify. Spotify runs on Android, Windows, Linux, gaming consoles, and smart TVs. Apple Music works best inside Apple’s walled garden. Spotify has roughly six times Apple Music’s user base. Cross-platform availability was a structural advantage from the first day.
VLC. VLC reverse-engineered support for every media format it had no legal right to play. It didn’t ask permission. It became the default media player for exactly that reason. When your tool plays everything, you stop needing any other tool.
The pattern is not subtle. The more flexible option wins. The locked-down option defends quality short-term and loses relevance long-term.
The harness landscape, May 2026
This section will be wrong in six months. That’s the point. Come back and check.
Claude Code (Anthropic) The current feature leader. The most mature harness, with the richest agentic capabilities on the market. It runs Anthropic models only. CLAUDE.md is its standards file format, but Claude Code does not read competitor config formats. If you’re comfortable in the Anthropic ecosystem and not worried about model optionality, it is genuinely the most capable harness available today.
GitHub Copilot CLI (Microsoft) Multi-model: Anthropic, OpenAI, Google, and xAI models are all available. Deep GitHub integration. Enterprise controls and reporting that satisfy most security and compliance requirements. It reads and follows CLAUDE.md. The system prompt philosophy keeps developers engaged in the workflow rather than waiting for output. Microsoft is actively closing the feature gap with Claude Code.
ChatGPT Codex (OpenAI) The hardest category to name. Codex is either a CLI with a UI or the lightest-weight IDE ever built, depending on how you look at it. It surfaces the things developers actually want from an IDE: diffs, file navigation, clickable interfaces. But it keeps the lightweight interaction model of a CLI rather than carrying the full weight of an editor. It imports your Claude Code work on setup, which is the interoperability move covered in the next section. The catch: once you’re in, you’re running OpenAI models. Same ecosystem lock-in as Claude Code, different ecosystem.
Cursor / xAI xAI’s position is that the IDE is the harness, not the CLI. Cursor is their current vehicle for that thesis. Strong IDE-native experience, model-flexible. The IDE-vs-CLI question is genuinely unresolved, and if xAI is right, every CLI-first harness has a ceiling.
OpenClaw and homebrew harnesses Early signal. OpenClaw and others are pushing the harness concept into shapes that corporate products haven’t explored yet. These are not production choices today, but they represent where community energy is pulling. Corporate equivalents are launching rapidly. The category is real.
Two tells
Copilot reads CLAUDE.md. Claude Code does not read Copilot’s equivalent.
One player is making it easy to bring your existing configuration with you. The other requires you to start fresh. Every example in the previous section has a version of this moment. Git reading your SVN history. Google Docs opening your Word files. VLC playing your proprietary formats. The player that says “bring your stuff, we’ll handle it” is playing offense. The player that says “start over with us” is playing defense.
ChatGPT Codex imports your Claude Code work.
When browsers started importing each other’s bookmarks, they were sending a signal: we are confident enough in our product to let you bring your history. We expect to win on merit, not on switching friction. Codex doing this with Claude Code work is the same move. It deliberately lowers the cost of trying it. That is not the behavior of a product that expects to win by trapping you.
Lock-in is a defensive strategy. The companies leaning into interoperability are playing offense.
What I decided, and why it doesn’t need to be your decision
My developers use GitHub Copilot CLI. My business users use Claude Desktop.
I picked Copilot CLI for developers because of GitHub integration (it fits how the team already works), multi-model availability (we are not locked to last quarter’s best model), enterprise controls and reporting (security and compliance requirements are real), and a system prompt philosophy that keeps developers engaged in the work rather than waiting for the AI to come back with something.
I am not telling you to do this. I am telling you that I made a reasonable call, it took less than a week to evaluate, and I am not losing sleep about whether it was the optimal decision. It was a good decision. Good enough is good enough when the space is moving this fast.
The choice you make today is not a ten-year commitment. The harness that wins in 2028 probably doesn’t exist yet. Make a call, stay positioned to adapt, and stop treating this like a decision that requires an oracle.
The prediction
The harness that gives you the most flexibility wins. Not specifically the one that reads everyone else’s config. Flexibility in the broader sense: flexibility to swap models, flexibility to update your workflow without rearchitecting everything, flexibility to stay current as the model layer moves.
Three months ago, Opus 4.6 was the consensus best coding model. Today, sentiment has shifted toward Codex 5.5. Six months from now, the leaderboard will have moved again. The organizations that will feel this least are the ones that chose a harness that treats model selection as a setting, not a commitment.
The model layer is moving faster than any organization can evaluate and re-decide. The harness that absorbs that movement without making it your problem is the structural winner.
Flexibility isn’t a nice-to-have. It’s the whole game.
Check back in six months.