Agent Intelligence & Skills — User Guide

Audience: Teams that want repeatable agent behaviour: always-on instructions, on-demand skills, guardrail rules, and lessons from incidents—across AI Chat, pipelines, and automation.

What this is

Agent Intelligence is the product area at /agent-intelligence where you configure how agents think and act in your workspace—without rewriting Python for every tweak.

Building block Role
Instructions Always-on behaviour text (tone, defaults, stack preferences).
Skills On-demand expertise: a trigger + full body loaded only when the agent calls read_skill(name).
Rules Hard constraints (“never drop production tables without approval”).
Lessons learned Documented mistakes and corrections so agents avoid repeating them.
Templates Preview and apply starter instructions, skills, or rules.

This is separate from the Knowledge Base (Part 2 — uploaded documents for RAG) and separate from Custom Agents & Tools (Part 1 — agent definitions and Python tools)—but all three work together.

Why skills matter

  • Context windows are finite. Dumping fifty long playbooks into every prompt wastes tokens and confuses the model.
  • Skills load selectively: the agent sees the skill trigger first; the full Markdown body loads when it invokes read_skill(name).
  • Operational evolution: add skills as new domains appear; tighten triggers when the wrong skill fires; pair with lessons after outages.

The five tabs (UI)

Tab Internal type What you configure
Instructions instructions Name, content (Markdown), priority (e.g. 0–100, higher first), optional agent filter (blank = all agents).
Skills skills Name, skill trigger (“when to use”), content, priority, optional agent filter.
Rules rules Name, content, priority, optional agent filter.
Lessons learned lessons Pattern (what went wrong), correct behaviour, category (see below).
Templates — Preview / Apply to spawn a new instruction, skill, or rule.

Lesson categories

general, planning, verification, code_quality, communication, architecture, testing — pick the category that helps retrieval and reporting.

Skills deep dive

Each skill has:

  • Name — stable identifier used with read_skill(name).
  • Skill trigger — short text the model sees before loading the full skill (when to consider this skill).
  • Content — full procedure, examples, SQL style, org glossary—Markdown.
  • Priority — ordering when multiple items compete.
  • Agents — optional restriction to specific agents only.

Troubleshooting skills

Problem Fix
Skill never loads Trigger too vague; align trigger with user phrasing.
Wrong skill loads Split skills; narrow triggers; add a rule to disambiguate.
Agent ignores a skill Check agent filter; raise priority; add an instruction anchor.

Instructions vs rules vs lessons

Type Use for
Instructions “How we work here” — stack, naming, doc style, default depth of answers.
Rules Non-negotiables — compliance, forbidden actions, data-handling.
Lessons Concrete “we tried X and it broke; do Y instead” with a category.

If behaviour is too loose, add rules. If it is too rigid, move nuance into instructions or skills.

Import, export, promotion

  • Export JSON — snapshot all instructions, skills, rules, lessons for backup or review.
  • Import JSON — merge into the workspace; UI reports counts imported.
  • Promote across environments — export from staging, import to production after review (same pattern as infrastructure config).

How this connects to the rest of Dagen

Feature Connection
AI Chat Execution modes (Guided / Semi / Auto) + Intelligence = safer automation.
Custom Agents & Tools Agents pick tools; Intelligence shapes when and how they use them.
Knowledge Base Files vs skills: files are reference material; skills are procedures.
Building Pipelines Fix with Agent benefits from lessons and rules about your stack.
Git Reviews Custom review instructions + Intelligence can align “what good SQL looks like” org-wide.

Route and access

  • Path: /agent-intelligence
  • Typically available to users who can manage workspace agent configuration (confirm with your admin if you do not see the menu).

Related

Last updated: February 21, 2026