🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
Argus by Perciqa
88% of AI agent pilots fail to reach production. The #1 reason is not bad models. It is that teams cannot prove their agents are reliable. Argus is the reliability layer every AI agent needs. | TRAJECTORY TRACING: Traditional observability gives you flat logs. Argus captures the full decision tree, every reasoning step, tool call, and model selection, as a structured replayable trace. When an agent breaks, you replay the trajectory and see exactly where reasoning went wrong. | INFERENCE FINOPS: Most teams know their GPU bill. Almost nobody knows their cost per successful agent task. Argus tracks token usage per agent and per task, flags runaway loops, and shows you exactly how much you save by routing to local self-hosted models versus cloud APIs. Local inference is always $0.00. | EVAL-IN-PRODUCTION: Argus runs LLM-as-judge evaluations on a configurable sample of production traffic. It detects accuracy drift before users do and generates a structured audit trail for every agent decision. | TECHNICAL APPROACH: No CUDA dependencies. ROCm-compatible by design. The demo agent uses hybrid model routing: simple queries go to Gemma running locally at $0, complex queries go to DeepSeek-V4 via Fireworks AI. Argus traces every routing decision and quantifies the cost split in real time. | STACK: Python SDK, FastAPI, SQLite, Next.js 16 dashboard, WebSockets, Fireworks AI, Docker. Open source, MIT licensed, self-hosted. github.com/perciqa/argus -- Built by Perciqa.
12 Jul 2026
