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 41-45 of 281 articles.

Product

Facio's Rollback Discipline: How AI Agents Undo Their Own Mistakes Without Making Things Worse

Jul 26, 2026Product

Facio's Rollback Discipline: How AI Agents Undo Their Own Mistakes Without Making Things Worse

AI agents take actions. Some actions are wrong: a customer is charged twice, an email goes to the wrong person, a deployment breaks production. The naive approach accepts the bad state and moves on. Facio's rollback discipline gives AI agents structured mechanisms to undo their own actions: reversible action design, automatic rollback triggers, compensating transactions for multi-step workflows, confirmation and verification, and learning loops that improve behavior. Bad state is reversed quickly; the system recovers cleanly.

Security

CaMeL and the Two-LLM Prompt Injection Defense: Why Most 2026 Deployments Still Fail the Trust Boundary That CaMeL Got Right

Jul 26, 2026Security

CaMeL and the Two-LLM Prompt Injection Defense: Why Most 2026 Deployments Still Fail the Trust Boundary That CaMeL Got Right

The CaMeL architecture separates the privileged LLM that processes trusted instructions from the non-privileged LLM that processes untrusted data, enforcing capability policies at the tool layer. Prompt injection surged 340% in 2026; most deployments still use a single LLM that cannot reliably distinguish instructions from data. The architectural commitment is the defense.

Human-in-the-loop

HITL and the Audit Trail of Effort: Why the Reviewer's Struggle Is the Most Honest Signal of Their Calibration

Jul 26, 2026Human-in-the-loop

HITL and the Audit Trail of Effort: Why the Reviewer's Struggle Is the Most Honest Signal of Their Calibration

The audit trail records decisions. The reasoning is the polished output. The struggle is hidden. But the struggle — the back-and-forth, the false starts, the abandoned reasoning, the rejected first drafts — is the most honest signal of the reviewer's calibration. A reviewer who struggles is engaging. A reviewer who doesn't struggle is rubber-stamping. Here is why the audit trail should capture the effort, not just the decision.

Engineering

MCP Spotlight: Time MCP Server — Anthropic's Reference Implementation for IANA Timezone Conversion, DST-Aware Math, and the DateTime-Default Reference for Agents

Jul 25, 2026Engineering

MCP Spotlight: Time MCP Server — Anthropic's Reference Implementation for IANA Timezone Conversion, DST-Aware Math, and the DateTime-Default Reference for Agents

The official Time MCP Server by Anthropic — 2 tools (get_current_time, convert_time) backed by the IANA tzdata database. DST-aware transitions, historical accuracy back to early 1900s, ~400 zones, alias support. The timezone-aware time-math default for AI agents in 2026. MIT-licensed, available via uvx or Docker.

Product

Facio's Tool Surface Discipline: How AI Agents Get Exactly the Capabilities They Need Without Exposing the Entire System

Jul 25, 2026Product

Facio's Tool Surface Discipline: How AI Agents Get Exactly the Capabilities They Need Without Exposing the Entire System

AI agents gain power through tools. The naive approach exposes every tool to every agent. The risk: an agent that's supposed to handle customer support questions can also trigger a payment refund; an agent compromised via prompt injection has the keys to the kingdom. Facio's tool surface discipline gives every agent a controlled capability set: per-agent allowlisting, argument scoping, context-aware dynamic loading, complete invocation logging, and anomaly detection. The agent is powerful enough to do its work; constrained enough to not be a liability.