Spotify shipped a meta-harness, and 36,000 sessions say it isn't a demo

Spotify's Portal team published Xirp on August 10: a vendor-neutral agentic development environment that manages concurrent sessions across Claude Code, Gemini CLI, and Codex, each session isolated in its own git worktree so dozens of agents can work the same codebase without stepping on each other.

The adoption number is what separates this from every other launch post. Thousands of Spotify engineers adopted it organically across more than 36,000 sessions. That's internal usage that happened before the public beta, which means the product survived contact with a real engineering org before it survived contact with Product Hunt. It debuted at #4 on Product Hunt's August 11 leaderboard with 116 upvotes.

The design decision I find most interesting: context is decoupled from any single harness, so your working state survives switching tools mid-project. Paired with Spotify Portal, session start pulls in component architecture, dependency graphs, ownership topology, and ADRs. That's the answer to a problem I hit constantly. Every harness has its own memory format, its own context assembly, its own idea of what your project is, and switching costs you everything you'd accumulated.

Now look at the rest of that same leaderboard. Bullet at #5 (113 upvotes). bb at #6 (101), orchestrating Claude Code, Codex, and OpenCode. Octomind Cloud and Hub at #9 (96), bundling 27 models behind one login with no API keys and per-second billing. Four of the top nine products in a single day are all selling the orchestration layer above coding agents rather than a coding agent.

Add VelaTerm open-sourcing its Tauri 2 terminal that runs nine coding agents as first-class sessions, and Ante shipping a 15MB single Rust binary that runs fully offline on pinned llama.cpp across 12+ providers. Different teams, different bets, same layer.

The pricing consequence is the part nobody's saying out loud. When the harness owns model selection, the model vendor loses the customer relationship. Octomind's 27-models-one-login pitch is explicitly that you shouldn't have an account with any of them. Every one of these products is a wedge between you and Anthropic or OpenAI, and the model becomes an interchangeable backend priced per token in a market where the buyer is a router optimizing on cost.

That's also the connection to Simon Willison's read on GitHub Models shutting down: agentic usage patterns made subsidized token access impossible, and free on-ramps are consolidating behind subscriptions. The harness layer is where that consolidation gets arbitraged. codex-bridge took 359 stars in two days doing exactly this, piping GPT-5 subagents into Claude Code through an existing Codex CLI OAuth session, no third billing relationship required.

If you're building a harness: portable context is the moat, not model coverage. Everyone will have every model within a quarter.