Zo Computer makes its AI agent 20 times more reliable with Vercel
Zo Computer adopted Vercel AI Gateway and SDK to get a single interface to models from OpenAI, Anthropic, MiniMax, GLM and Fireworks. The retry rate dropped from 7.5% to 0.34%, chat success rose to 99.93%, and latency fell by 38%. New models can now be added in seconds.
AI-processed from Vercel Blog; edited by Hamidun News
Zo Computer, a personal cloud platform with an AI agent, has made its system 20 times more reliable. The key was adopting Vercel AI Gateway and SDK, which unified how it works with different model providers.
The problem: one adapter per API
Zo Computer lets users work with any AI model and upload their own keys. That means the platform has to support OpenAI, Anthropic, MiniMax, GLM, Fireworks and many others. Previously, each provider required its own adapter, with different image handling, different key-management logic and different edge cases. When a new model shipped every week, an engineer had to write an adapter by hand, test it and deploy it. For a team of 8 people, this was a constant drain on time.
The starting metrics were grim: a 98% success rate with a 7.5% retry rate. Every 50th user message either failed or needed a retry.
The solution: SDK +
Gateway instead of a hundred integrations
Zo moved to Vercel AI SDK and AI Gateway, two layers that split the responsibilities.
AI SDK replaced all the hand-rolled adapter code. Instead of a separate implementation for each provider, engineers got a single interface. Image handling is normalized and response formats are unified.
AI Gateway took over the infrastructure layer:
- Automatic retry logic instead of manual management
- Routing between providers and fallback to alternative models
- Real-time monitoring and optimization
- New models are added through config, with no code to write
After the switch, Zo could add support for new models in less than a minute.
The numbers: a 20x improvement and more
The results are impressive:
- Retry rate: 7.5% → 0.34% (a 20x reduction)
- Success rate: 98% → 99.93%
- P99 latency: 131 sec → 81 sec (down 38%)
- Time to add model: <1 minute, no code
Rob Cheung, co-founder of Zo Computer: "We didn't even suspect the problem until we moved to Vercel. The numbers showed the truth: there were far too many failures before."
What this means
For startups like Zo, this means they no longer depend on in-house infrastructure built around each model. The team can focus on the product, on how users interact with the AI agent. Zo is aiming for a million new users in 2026, and the technology now makes that possible.
- Nov 1, 2026Local models lose to Opus and Codex: 0.21 on a MacBook versus 1.00 for Codex
- Nov 1, 2026Adobe's EditAgent judge picks a usable edit 81.3% of the time, versus 74.8% for EditScore
- Nov 1, 202616 models tested on 7 search tools: Exa averaged 10.5 out of 14, SearXNG about 3
- Nov 1, 2026Google's StreamVibe video agent beats other models by 10.7% on response timing and quality
Key Questions Answered
What are the key details on The problem: one adapter per API?
Zo Computer lets users work with any AI model and upload their own keys. That means the platform has to support OpenAI, Anthropic, MiniMax, GLM, Fireworks and many others. Previously, each provider required its own adapter, with different image handling, different key-management logic and different edge cases. When a new model shipped every week, an engineer had to write an adapter by hand, test it and deploy it. For a team of 8 people, this was a constant drain on time. The starting metrics were grim: a 98% success rate with a 7.5% retry rate. Every 50th user message eithe
What are the key details on The solution: SDK + Gateway instead of a hundred integrations?
Zo moved to Vercel AI SDK and AI Gateway, two layers that split the responsibilities. AI SDK replaced all the hand-rolled adapter code. Instead of a separate implementation for each provider, engineers got a single interface. Image handling is normalized and response formats are unified. AI Gateway took over the infrastructure layer: Automatic retry logic instead of manual management Routing between providers and fallback to alternative models Real-time monitoring and optimization New models are added through config, with no code to write After the switch, Zo could add supp
What are the key details on The numbers: a 20x improvement and more?
The results are impressive: Retry rate: 7.5% → 0.34% (a 20x reduction) Success rate: 98% → 99.93% P99 latency: 131 sec → 81 sec (down 38%) Time to add model: <1 minute, no code Rob Cheung, co-founder of Zo Computer: "We didn't even suspect the problem until we moved to Vercel. The numbers showed the truth: there were far too many failures before."
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.