Spine AI, a YC Summer 2023 company, launched Spine Swarm on March 13, 2026—a multi-agent system that uses an infinite visual canvas to complete complex non-coding projects. The system achieved top rankings on two Google DeepMind benchmarks, scoring 87.6% on DeepSearchQA and claiming the #1 position on GAIA Level 3.
8.1% Lead Over Perplexity on DeepSearchQA
Spine Swarm scored 87.6% on the full 900-question DeepSearchQA dataset, positioning it 8.1 percentage points ahead of Perplexity and 21.5 points ahead of Google's own Gemini Deep Research. The benchmarks tested with zero human intervention, with agents running fully independently without clarification questions.
Founders Ashwin and Akshay, who have been friends for 13 years and took their first ML course together at NTU, named the company after the North Spine area where they studied. After going through YC in 2023, they spent three years iterating on the product architecture.
Visual Canvas Architecture Enables Parallel Agent Work
The core architecture uses a visual canvas with specialized blocks for different capabilities: LLM calls, image generation, web browsing, apps, slides, and spreadsheets. Each block represents a specialized unit that can pass context to connected blocks, enabling multiple agents to work in parallel when subtasks lack dependencies.
A central orchestrator decomposes tasks into subtasks and delegates to specialized persona agents. The system dynamically selects from 300+ models for each subtask, allowing a single workflow to use different providers—OpenAI for one step, Nano Banana Pro for image generation, and Claude for app generation.
Human-in-the-Loop Without Workflow Reruns
Agents can pause to request clarification before continuing, and users can select canvas blocks to iterate through chat without rerunning entire workflows. The founders argue that "chat is the wrong interface for complex AI work" because it creates a linear thread while real projects are non-linear. The canvas provides persistent, structured representation that prevents context degradation as work passes between agents.
Use cases include SEO analysis, competitive landscape research, fundraising pitch decks with financial models, feature prototyping from screenshots or PRDs, and deep-dive learning plans. The system operates on a usage-based credit model tied to block usage and underlying models, with a free tier available.
Key Takeaways
- Spine Swarm achieved 87.6% on DeepSearchQA, ranking #1 and outperforming Perplexity by 8.1 percentage points
- Visual canvas architecture enables parallel multi-agent work with specialized blocks for different capabilities
- System dynamically selects from 300+ models for each subtask, using different providers within single workflows
- Agents run with zero human intervention on benchmarks but support human-in-the-loop clarification for production use
- Founded by YC S23 alumni Ashwin and Akshay after 3 years of product iteration on multi-agent systems