Glossary

Core Concepts

Agent

An autonomous software system that perceives its environment, makes decisions, and takes actions to achieve goals without explicit instruction for every step. Can operate independently and continuously.

Orchestrator / CEO Agent

A specialized agent that receives high-level goals and delegates specific tasks to specialist agents. Responsible for breaking down complex work, routing to the right specialists, and verifying results. Enables scalable multi-agent systems.

Sub-agent

An agent that operates under the direction of an orchestrator or parent agent. Focused on specific domains or tasks (e.g., “customer support specialist,” “data analyst”). Reports outcomes back to the orchestrator.

Delegation

The process of an orchestrator agent breaking a goal into subtasks and assigning them to specialized sub-agents. Critical for building scalable systems without overloading any single agent.

Architecture & Design

Agent Tree Architecture

A hierarchical structure where one orchestrator agent delegates to multiple specialized sub-agents, which may themselves delegate to other sub-agents. Creates efficient, maintainable, scalable systems. Also called “multi-agent orchestration.”

Verification Loop

A pattern where output from one agent (or human) is checked by another agent or human before being finalized or sent to stakeholders. Ensures quality and catches errors before they cause problems.

Smart Model Routing

The practice of selecting the right LLM model for each specific task. For example: use Haiku for high-volume data processing, Sonnet for balanced tasks, Opus only for complex reasoning. Optimizes cost and speed.

Memory System (3-Layer)

A structured approach to agent memory with three levels:

SOUL.md — Core identity, values, and decision-making principles. Rarely changes.
USER.md — Information about the human/context the agent is helping. Used in every session.
MEMORY.md — Long-term learnings, important decisions, and recurring patterns. Reviewed and updated periodically.

Also includes daily logs (memory/YYYY-MM-DD.md) for raw session notes.

OpenClaw Specifics

OpenClaw

An open-source framework for building AI agents designed for the mid-market company use case. Provides tooling for delegation, memory management, model routing, and verification loops. Built with the belief that agents should be autonomous but verifiable.

SOUL.md

A file that captures an agent’s core identity, values, and principles. What the agent “believes” about its role. Informs decision-making and helps maintain consistency across sessions.

USER.md

A file containing context about the human the agent is helping. Includes preferences, timezone, ongoing projects, and communication style. Loaded at the start of each session for personalization.

MEMORY.md

A curated, long-term memory file. Contains important decisions, insights, and learnings worth retaining beyond daily sessions. Updated periodically (every few days) by reviewing daily logs and extracting key patterns.

Heartbeat

A periodic check (typically every 30-45 minutes) where an agent reviews pending tasks, checks notifications, and takes proactive action if needed. Different from cron — uses conversational context and can batch multiple checks.

Example heartbeat checks: email inbox, calendar events for the next 24h, mentions on social, weather (if relevant).

Cron Job

A scheduled task that runs at a specific time (e.g., “9:00 AM every Monday”). Isolated from main session context. Better for: exact timing, one-off reminders, or tasks that should run independent of ongoing conversations.

Technical Terms

Token

A unit of text (roughly 4 characters) used to measure LLM input/output costs and model capacity. Larger responses use more tokens. Token count affects both cost and response time.

Context Window

The maximum number of tokens an LLM can read at once. For example:

Claude Haiku: 200K tokens
Claude Sonnet: 200K tokens
Claude Opus: 200K tokens

Larger context windows enable agents to remember longer conversations and work with bigger documents.

Model Tier

The category/performance level of an LLM:

Haiku — Fast, cheap. Best for high-volume, simple tasks.
Sonnet — Balanced. Best for most agent work.
Opus — Slow, expensive, powerful. Best for complex reasoning and planning.

Choosing the right tier for each task drives down costs significantly.

Advanced Patterns

ARD (Agent Requirement Document)

A structured template for defining what an agent should do. Similar to a software spec, but focused on behavior, constraints, and edge cases. Helps catch design issues before building.

See Agent Requirement Document for the full template.

Delegation Framework

A set of patterns for how orchestrators should break down work and assign it to sub-agents. Includes:

How to specify a subtask clearly
How to handle failures or edge cases
How to verify work is complete
How to learn which specialists are reliable

See Delegation Framework.

Concepts & Philosophy

High-Leverage Agent (Thesis)

AI agents provide maximum ROI at mid-market companies because:

Complex enough to need sophisticated operations (finance, HR, sales, customer success)
Small enough that change is still possible (no legacy systems bureaucracy)
Right-sized for a small team of agents to eliminate significant administrative work

Smaller companies don’t have the operational complexity. Larger companies have entrenched systems and slower decision-making.

Autonomy vs. Oversight Trade-off

A core tension in agent design:

More autonomy = faster, 24/7, requires less human attention
More oversight = safer, catches errors, but slower and more labor-intensive

The right balance depends on risk tolerance. A scheduling agent can be fully autonomous. A financial decision agent needs verification loops.

See something missing? Open an issue on GitHub with glossary suggestions.