Why Open Source Will (Still) Win in the Age of Agents
Closed platforms get sandwiched; open is the right call.
Last week, Cursor announced a probable acquisition by SpaceX. A few months ago, Windsurf got picked apart by frontier labs. But this shouldn’t be surprising. Agent harnesses ending up yoked to particular labs is downstream of the structural pressures of running a closed AI platform.
Even if you thought you were just choosing a coding tool with Cursor or Windsurf, it turns out you were also choosing a particular lab’s models. For coding harnesses, maybe that’s tolerable – switching costs are low. But for operational agents inside your business, the same dynamic will be far harder to unwind.
Closed harnesses aren’t competing with the frontier labs, they’re auditioning for them. The valuable user data that they once treated as a moat is actually bait. And if you want a way to run agents that won’t leave you stuck to one model family, it has to be open source.
Closed platforms are being sandwiched by the market.
A closed harness sits between the model below and the workflow above. Its only differentiation is doing what the lab does, faster or slightly better. Cursor was treating its accumulated user data as its moat. As all the labs move downstream from model to product, the harness has nothing to defend with.
Independent closed harnesses like Cursor and Windsurf start independent and get absorbed. Lab-native harnesses like Claude Code and Codex are absorbed from the get-go. Either way, your tooling ends up tied to one lab’s roadmap, pricing, and technical choices.
You might object that being tied to one lab is fine when that lab is the leader. But as we’ve written about over and over, the one predictable technological outcome is that model leadership has flipped repeatedly in the last 24 months and will flip again. “The best model right now” is not a durable bet.
Open source breaks vendor lockin in every generation.
Every generation of computing infrastructure runs the same fight for business customers. Mainframe vendors shipped fully closed stacks: hardware, OS, applications, all locked together. Unix opened the operating system layer and a third party application market was born, but Unix licenses became the new lock-in. Linux broke that, and Linux now runs almost everything that isn’t a desktop.
Each cycle, proprietary wins early business adopters. Open source wins the long game. Each cycle, buyers who held the line on portability stayed flexible. Buyers who optimized for whatever was best right now paid switching costs later, sometimes catastrophic ones.
Large language models are the new infrastructure layer. Agentic frameworks are the new middleware. The same fight is on, and the buyer’s job is the same as it has always been: don’t get locked in.
The “open source isn’t ready yet” excuse is gone.
In every prior cycle, closed platforms had a window. The proprietary stack was genuinely better for a few years before open alternatives caught up. That window is what made “we’ll worry about portability later” a defensible position.
For agentic harnesses, that pattern has reversed and it was an open-source framework that hit escape velocity for always-on agents. When OpenClaw launched in November 2025, it was an instant hit, and Anthropic and OpenAI both scrambled to add OpenClaw-like functionality to their products.
And the OpenClaw stats back up the feeling. In barely five months of existence OpenClaw had accumulated: 360,000 GitHub stars, 74,000 forks, 1,800 contributors, and 44,000 community-built skills on ClawHub.
The point isn’t that OpenClaw specifically will win as the enterprise agentic harness of choice. It is clear that the modular, open agentic stack is here now, not in two years. The grace period closed platforms used to acquire users is over before it started.
Openness is the best defense to lock-in.
Build internal expertise on open agentic frameworks now. Not “evaluate,” not “track.” Get hands on. Your team’s familiarity with the open stack is what makes portability real. Without it, “we can switch later” is a story you tell yourself.
Run operational agents on open frameworks, not closed platforms. The further into your business an agent runs, the more painful it is to rip out. Customer-facing workflows, internal ops, engineering automation. These are the places where lock-in compounds. Start them on open foundations.
This is the strategic bet we’re making with KiloClaw. A hosting business built on an open framework doesn’t need to hold you captive to stay viable — it has to earn your continued business on quality and price, the same as any commodity infrastructure provider. It puts the onus on us to be more valuable than exporting your configuration and running OpenClaw yourself – and we think that’s a good deal. We win by being useful, giving users and teams the features they need in an agentic engineering platform.
Choices you won’t regret when things move fast
I don’t know where Anthropic, OpenAI, or SpaceX will be in two years, and neither does anyone else. So the move isn’t to pick the winner. It’s to make sure being wrong is cheap.
The incredible velocity of the market for AI can make it hard to see that vendor lock-in dynamics are playing out the same as always. And, like always, open source is the most reliable hedge for vendor lock-in.
Agents don’t change that dynamic. Instead, they raise the stakes because the deeper an agent runs inside your business the more it costs to be wrong about who holds the keys.