Vercel found a simple filesystem beat their multi-agent system — plus 19 OSS agent tools
MaryamMiradi · x · 2026-10-15
A rundown of 19 GitHub repos for AI agent building in 2026, organized around the six architecture stages Vercel presented at AI Engineer:
- Stage 1, one big prompt: Vercel AI SDK, LiteLLM, Instructor — first test valid SQL generation.
- Stage 2, multi-agent: LangGraph, CrewAI, Google ADK, OpenAI Agents SDK — specialization but lost context in handoffs.
- Stage 3, one stateful agent: PydanticAI, Deep Agents, Mastra — plan, execute, recover from errors.
- Stage 4, filesystem agent: Claude Code, just-bash, Vercel Sandbox — file system + Bash roughly doubled the eval score.
- Stage 5, reusable skills: Anthropic Skills, Skills CLI, Vercel Agent Skills.
- Stage 6, production: Eve, Workflow DevKit, OpenTelemetry.
Key takeaway: better context, familiar tools, and a simple file system can outperform elaborate multi-agent workflows; the trick is knowing when to add or remove complexity. The author plans to benchmark LangGraph vs Google ADK vs PydanticAI vs Deep Agents vs Eve next.
More from coding & agent
- Running 40+ Local Agents Turns Your Machine Into a Heater: Bottleneck Is I/O, Not Agent Design — rachittshah · 2026-10-15
- Enroute open-sources Git server hitting 7,600 concurrent pushes/s per repo on Raft and S3 — DanielLockyer · 2026-10-15
- INT21 generates Rust inference runtime for Drama 3 in 1 hour, 63% faster than SGLang — bingxu_ · 2026-10-15
- Claude Platform adds OpenTelemetry tracing for Messages API and Managed Agents sessions — ClaudeDevs · 2026-10-15
- Build a Code Graph Before Agents Read a Single File, Says Limestone — alex_verem · 2026-10-15
- Hermes Phone Harness plugin gives the agent real iPhone and Android control — lifebypixels · 2026-10-15