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 46-50 of 281 articles.

Human-in-the-loop

HITL and the Hesitation Signal: Why the Reviewer's Pause Before Clicking Is the Most Valuable Information They Never Log

Jul 25, 2026Human-in-the-loop

HITL and the Hesitation Signal: Why the Reviewer's Pause Before Clicking Is the Most Valuable Information They Never Log

The reviewer's hesitation — the moment of pause before clicking approve — is the most valuable signal in HITL. It contains the reviewer's doubt, their pattern recognition, their gut feeling, their risk assessment. The system captures the click but not the pause. The pause is invisible. The pause is what separates the accurate reviewer from the rubber stamp. Here is why the hesitation matters, how to measure it, and what changes when the system finally sees it.

Engineering

MCP Spotlight: PostgreSQL MCP Server — The Database Bridge With Read-Only Default, Query Allow-List, Parameterized Statements, and the Schema-Aware Default for Agents

Jul 24, 2026Engineering

MCP Spotlight: PostgreSQL MCP Server — The Database Bridge With Read-Only Default, Query Allow-List, Parameterized Statements, and the Schema-Aware Default for Agents

The official PostgreSQL MCP Server by Anthropic — ~12 focused tools (list_schemas, list_tables, describe_table, query, explain). Read-only is the default, write-mode requires explicit opt-in with statement-type filtering. Parameterized queries via $1, $2 placeholders (SQL injection impossible). MIT-licensed.

Product

Facio's Memory Hierarchy Discipline: How AI Agents Decide What to Remember, What to Forget, and What to Surface at the Right Moment

Jul 24, 2026Product

Facio's Memory Hierarchy Discipline: How AI Agents Decide What to Remember, What to Forget, and What to Surface at the Right Moment

AI agents accumulate state: conversation history, tool results, intermediate reasoning, user preferences, learned facts. The naive approach stores everything and hopes the model sorts it out. The disciplined approach treats memory as a hierarchy with explicit retention policies, decay functions, relevance scoring, and surfacing mechanisms. Facio's memory hierarchy discipline gives AI agents the structural framework to remember what matters, forget what doesn't, and surface the right information at the right moment without overwhelming the context window.

Security

AIBOM and CycloneDX ML-BOM: Why AI Agent Supply Chains Need Their Own Software Bill of Materials in 2026

Jul 24, 2026Security

AIBOM and CycloneDX ML-BOM: Why AI Agent Supply Chains Need Their Own Software Bill of Materials in 2026

The May 2026 G7 AI SBOM Framework and the HuggingFace pickle RCE wave made AIBOM mandatory. CycloneDX ML-BOM is the format, Sigstore/Cosign and in-toto are the provenance, the ingestion gate is the enforcement, and the five-stage maturity ladder is the roadmap. Five supply chain layers (model weights, tokenizers, frameworks, tools, data sources) require their own AIBOM components.

Human-in-the-loop

HITL and the Stop Rule: Why Every Reviewer Should Have a Personal Threshold for Rejecting on Sight

Jul 24, 2026Human-in-the-loop

HITL and the Stop Rule: Why Every Reviewer Should Have a Personal Threshold for Rejecting on Sight

Every experienced reviewer has a "stop rule" — a personal threshold below which they reject the action immediately, regardless of the policy. The stop rule is not in the manifest. It's not in the training. It's not in the metrics. It's in the reviewer's intuition. The stop rule is the most underrated signal in HITL — and the most invisible. Here is why personal stop rules matter, how to surface them, and what changes when the system's rules and the reviewer's rules are aligned.