Hey Ph! We created Foil because we realized we were spending more time supervising our AI agents than actually building them. We had signs, dashboards, logs, but understanding them all was a full-time job. We were the bottleneck, manually reviewing thousands of traces to determine when and why agents were trapped, decoys, etc.
So we asked: What if we reimagined the way we monitor AI agents?
A foil is basically an AI agent whose job it is to watch your other agents. It learns how each one behaves (we call these agent profiles), sets its own health checks (anchors), and contextualizes problems not only with “increased error rates” but why and what has changed. A customer support bot and a code review agent are evaluated differently because they should be.
Here’s what’s under the hood:
👁️ Agent Profiles – Foil automatically learns each agent’s behavior: toll patterns, error rates, traffic patterns. A living baseline, not a static dashboard.
🎯 Anchors – automatically generated health checks (eg “defect rate < 5%)" that evaluate each mark against the profile
🔎 Smart Search — Query all your agents, brands, customers and models in natural language. Ask “Which agent has the highest error rate?” And get quick answers with charts.
👣 Tracing – Understand every decision your agent makes.
🔍 Detection – Deception, behavior escalation, loops, prompt injection, PII leakage, and RAG grounding failures – caught in real time
📊 Metrics – Dashboards per agent for cost, latency, quality and usage over time
🧠 Feedback loop – flag false positives or confirm real problems. Foil learns from you and gets more accurate over time.
🎇 Multimodal Support – We support all file types for documents, images, audio, video, code used for agent training and semantic search.
🎁 Product Hunt Exclusive – 50% off Pro for 3 months with code FOILPH
Try this: run npx @getfoil/foil-js wizard –dir
Ask us anything here!
Quick start example: github.com/getfoil/foil-examples