Facio Blog
Payload-powered product notes, security writing, HITL patterns, and operational guidance from the Facio runtime: long sessions, Placet approvals, audit trails, memory, providers, channels, tools, and Docker-first operations.
Showing 181-185 of 235 articles.
Human-in-the-loop
Most teams can't answer a simple question: "Is your HITL system actually working?" They track approval counts and respond times — vanity metrics that say nothing about oversight quality. Here are the 7 metrics that actually measure HITL effectiveness.
Engineering
DecisionNode turns architecture decisions into vector embeddings searchable by AI agents over MCP. 9 tools, free Gemini embedding tier, a force-directed graph UI, and live pulse showing which decisions your AI queries in real time.
Product
Connecting an MCP server is step one. Managing it in production — toggling it on and off, editing its configuration, restarting it after a crash, routing it to different HTTP transports — is where most platforms stop. Facio's manage_mcp tool gives agents full lifecycle control over their own tool infrastructure, with HITL gating on every destructive change.
Security
Standard Docker containers share the host kernel — and 1 in 8 AI breaches now involves an agentic system. From Cohere's CVSS 9.3 sandbox escape to Claude Code's configuration injection, the evidence is in: containers aren't sandboxes. Here's what isolation actually requires in 2026.
Human-in-the-loop
Prompt injection attackers don't need to hack your infrastructure. They just need to convince your agent to skip the approval gate. Research shows prompt-based HITL fails 74.6% of the time against adversarial input. Here's why — and how deterministic pre-action enforcement stops it at 0%.