SwarmTrace — pytest for AI Agent Swarms

SwarmTrace is an open-source developer toolkit that brings pytest-style observability to multi-agent LLM systems. Built at AMD Hackathon 2026, it solves one of the biggest pain points in production AI: you can't fix what you can't see. With a single @observe decorator, SwarmTrace automatically traces every agent call in a swarm — capturing latency, token usage, cost, and execution order in a beautiful nested tree. It works with any LLM provider, runs fully offline, and requires zero infrastructure setup. Key features include: - Multi-Agent Tracing — visualize your full agent tree with parent/child relationships, timing, and token costs per call. - Token Budget Manager — set hard token limits per agent with configurable warnings, so no single agent silently burns through your budget. - Tool Attention (arXiv:2604.21816) — a semantic similarity engine that selects only the most relevant tools for each query, reducing tool token overhead by up to 95%. - Regression Detection — compare two prompt versions across a test set and automatically flag semantic regressions before they reach production. - Web Scraping Tracing — trace live web scraping inside agent pipelines with full observability. - Async & Thread-Safe — full support for asyncio-based parallel agent swarms. - CLI Dashboard — browse, replay, and export traces from the terminal using rich color output. Benchmarked on AMD Instinct MI300X (192GB VRAM) via AMD Developer Cloud, SwarmTrace adds less than 1ms of overhead per agent call across swarms of 20+ agents. It is fully open-source (MIT), free forever, and self-hosted — making it a production-ready alternative to paid tools like LangSmith.