Agent surface bifurcation — three distinct moats, three different races
Gemini Spark ships personal agents to consumers. Cursor 2.5 ships parallel sub-agents to IDEs. Windsurf 2.0 ships autonomous cloud agents bundled with Devin. Three product categories, three different moats, three different races. The 'agent market' is becoming three markets.
The same week, three different products
Three releases:
- Gemini Spark — Google's 24/7 personal AI agent reasoning across Gmail, Calendar, Drive, Photos, YouTube. Consumer-tier, beta to Ultra subscribers.
- Cursor 2.5 Build in Parallel — concurrent sub-agent execution on the same git working tree, reviewer agent reconciles. IDE-tier.
- Windsurf 2.0 Cascade + bundled Devin Cloud and Devin Terminal CLI. IDE-plus-cloud-agent tier.
All three are 'agent products.' The category overlap stops there. Spark and Cursor 2.5 don't compete; Spark and Windsurf don't compete; Cursor and Windsurf compete narrowly on in-IDE orchestration. The market is three markets stacked under one word.
Three distinct moats
| Surface | Moat | Lock-in |
|---|---|---|
| Consumer personal agent | Cross-app data graph | User's email, calendar, photos |
| In-IDE concurrent orchestration | Workflow state + team collaboration | Engineer's daily editor habits |
| Autonomous cloud engineer | Long-running task delegation + Cloud Spaces | Build pipelines, deployed environments |
Each lane has a different switching cost. Consumer personal-agent switching cost is the user's data graph; IDE switching cost is the engineer's muscle memory plus team plugins; cloud-engineer switching cost is the integrated build-and-deploy environment.
Where the model layer sits in this
The underlying model is becoming the commodity input across all three lanes. Cursor 2.5 routes to Claude and GPT family. Windsurf 2.0 routes to Devin's model-agnostic backend. Gemini Spark uses Gemini. The model-lab tier wins or loses depending on whether the orchestration vendor picks its model — not on whether the orchestration vendor switches to it.
The orchestration vendors are the new buyers. The labs sell to them, not to end users.
The race within each lane
- Consumer personal agent. Gemini Spark vs OpenAI Operator vs Anthropic Claude Cowork vs (eventually) Apple Intelligence's evolution. Q3 2026 watch: cross-app task completion rate benchmarks.
- In-IDE concurrent orchestration. Cursor 2.5 vs Windsurf 2.0 vs (eventually) GitHub Copilot's full MCP migration. Q3 watch: engineering-team installed-base data.
- Autonomous cloud engineer. Devin (Cognition) vs Replit Agent vs (emerging) Claude Code's cloud-tier ambitions. Q3 watch: long-running task success rate metrics.
The honest read is that all three races have at least 18 more months to run. Procurement teams should stop trying to pick a single 'agent winner' and instead pick one vendor per lane, with explicit migration plans if the lane's leader changes.
CNBC — Gemini Spark → · Lushbinary — AI coding agents 2026 → · AIToolTier — Windsurf 2.0 review →