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 36-40 of 81 articles.
Engineering
Ejentum exposes four cognitive harnesses (reasoning, code, anti-deception, memory) as MCP tools that inject structured scaffolds into LLM context — catching sycophancy, hallucination, and causal shortcuts before they reach users. Benchmarked at +10.1pp across 180 tasks.
Product
Most agent platforms treat human approval as a binary step: approve or reject. Facio's HITL tools — ask_approval, ask_form, ask_selection, and ask_text_input — turn human-agent interaction into structured workflows with custom options, password-gated credential capture, markdown review, and timeout handling. Here's how each tool works and when to use it in production.
Security
Memory poisoning turns prompt injection into a stateful attack. MINJA achieves 95% injection success, AgentPoison exceeds 80% at under 0.1% poison rate, and Gemini's memory feature was compromised through delayed tool invocation. Here's what defense in depth actually requires.
Human-in-the-loop
Every HITL system eventually hits the same operational question: the reviewer didn't respond. Does the agent wait forever? Auto-approve? Auto-deny? The answer isn't a setting — it's a timeout strategy with consequences for safety, throughput, and compliance.
Engineering
The official WebDriverIO MCP server gives AI agents 34 tools to automate Chrome, Firefox, Edge, and Safari browsers — plus iOS and Android apps via Appium — all through a unified interface. Session recording, device emulation, and BrowserStack support included.