Facio Blog

Practical notes on human-reviewed AI agents.

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 1-5 of 187 articles.

Engineering

MCP Security Architecture 2026: Threat Models, Sandboxing, and the Defense-in-Depth Layers Every Production Agent Deployment Needs

The defense-in-depth reference architecture every team deploying MCP in production needs. Six layers: server verification → process sandboxing → credential isolation → per-tool RBAC → HITL approval → audit trail. Plus prompt-injection defense and multi-agent trust boundaries. Companion to the MCP Spotlight series.

Jul 11, 2026Engineering

Read article

Product

Facio's Cascading Timeouts: How AI Agents Bound Latency at Every Layer of the Stack

Jul 11, 2026Product

Facio's Cascading Timeouts: How AI Agents Bound Latency at Every Layer of the Stack

An AI agent's response time is the sum of many operations: model inference, tool calls, database queries, API requests, HITL pauses. Each operation has its own latency. Each operation can fail by being slow. Without discipline, a single slow operation makes the whole agent slow. Facio's cascading timeouts give teams the structural discipline to bound latency at every layer of the stack. The agent has time budget per operation; the operations chain into a budget per session; the session fails fast rather than hanging.

Security

When One Compromised Agent Becomes a Hundred Compromised Systems: The Lateral Movement Architecture You Cannot Detect at the Network Layer

Jul 11, 2026Security

When One Compromised Agent Becomes a Hundred Compromised Systems: The Lateral Movement Architecture You Cannot Detect at the Network Layer

A single compromised AI coding agent pivoted through the organization's npm registry, reached the CI/CD system, and propagated a malicious package to twelve production services. The propagation was authorized; the tool calls were legitimate; the only signal was the agent's reasoning. Lateral movement through AI agents is structural, not solvable by traditional tooling.

Human-in-the-loop

HITL and the Cost of Saying No: Why Reviewer Rejection Is the Most Expensive Decision and How to Make It Worth It

Jul 11, 2026Human-in-the-loop

HITL and the Cost of Saying No: Why Reviewer Rejection Is the Most Expensive Decision and How to Make It Worth It

Most HITL systems optimize approval velocity — actions approved per hour. But the most expensive decision the reviewer can make is rejection. Rejection costs the system the work the agent did, the latency the customer waited, the customer experience of the failed action, the cost of the corrective action. The system that doesn't value rejection is the system that loses the value of the most important review decision.

Engineering

MCP Spotlight: Memory MCP Server — Anthropic's Reference Implementation for Persistent, Shareable, Graph-Based Agent Memory Across Conversations

Jul 10, 2026Engineering

MCP Spotlight: Memory MCP Server — Anthropic's Reference Implementation for Persistent, Shareable, Graph-Based Agent Memory Across Conversations

The official Memory MCP Server by Anthropic — 9 tools built around a graph-based knowledge model with three primitives: entities (nodes), relations (typed edges), observations (facts). JSONL storage that's persistent across sessions, version-controllable with git, and shareable across agents. MIT-licensed.