Work Desks and Specialty Agents: The Architecture Between Sessions and Teams
Sessions, jobs, work desks, and specialty agents are four names for the same primitive — and knowing which one you mean changes how you design, debug, and gover
Field Manual · Working Notes
Triadive: a field manual for working with AI agents
Principles for the human / agent / robot triad — written at the level of concepts so they outlive any platform, project, or year. Concepts, workflows, and lessons drawn from the practice of running long-lived agents in OpenClaw.
Latest from the manual
- 2026 · 08 · 29DISPATCHESDispatch #5 — The CEO-agent bottleneck: why centralized review is the operator's most expensive habit
- 2026 · 08 · 26CONCEPTSFrom Prompts to Agents: The Gap You're Probably Skipping
- 2026 · 08 · 26CONCEPTSFrom Prompts to Desks: Why One Agent Is Not Enough
- 2026 · 08 · 26CONCEPTSWhat Is an LLM?
- 2026 · 08 · 26WORKFLOWSSetting Up Your First Three Desks
- 2026 · 08 · 26WORKFLOWSSetting Up Your First Three Work Desks: A Practical Starting Point
Reading paths
Three guided routes through the manual: start here, memory & goals, and patterns & lessons.
Start here
The three pieces to read in the first 30 minutes. Agent loop, the loop-vs-graph distinction, and how the runtime surfaces you use every day fit together.
What Is an Agent Loop?
An agent loop is the repeating cycle an AI agent uses to make progress on a goal.
Loops vs Graphs: When to Stop Iterating and Start Routing
Once you have built one agent loop, the temptation is to make it bigger.
OpenClaw Control on Mac and in the Browser
OpenClaw gives you two useful ways to work with the same system: the macOS app and the browser-based Control UI.
Memory & goals
The durable layers. How agents remember, why long-running agents forget, and how an explicit goal turns an agent into something an operator can follow.
Wiki Memory: The Fourth Layer Most Agent Systems Skip
The three-layer memory model — short-term, long-term, semantic — is the floor.
Why memory is the hardest part of any long-running AI workflow
The thing that breaks first in a long-running AI workflow is almost never the model.
Pursuing Goal in OpenClaw
OpenClaw's Pursuing goal feature is small on the surface and important underneath.
Patterns & lessons
Three patterns that survive contact with real work. What worked, what broke, and what to keep in your back pocket.
Workflow Review: The Daily Check-in Loop
Most workspaces need a daily "what changed?" check.
Lessons Learned: Why Agents Stall Without Clear Termination
Across long-running agents, the failures that cost the most are not reasoning failures.
Lessons Learned: Memory Design Mistakes in Early Agent Work
Two ways to think about agent memory. The first: the agent writes notes, the user reads them when something goes wrong.
Glossary
The vocabulary every Triadive article assumes you know. A separate reference layer — read these once, then refer back as needed.
From the editor
Dispatch #5 — The CEO-agent bottleneck: why centralized review is the operator's most expensive habit
Adding more desks does not make a multi-agent system faster if every desk still reports to the same overloaded coordinator. The CEO-agent bottleneck is the pattern that breaks most desk-based architectures, and it is almost entirely self-inflicted.
About this manual
Triadive is a working field manual for the human / agent / robot triad. The examples lean on OpenClaw because that's where the practice is most visible right now — but every principle here is meant to outlive the platform it came from.
- What this manual is. Principles-first. Concepts, workflows, and lessons drawn from real operator practice.
- What this manual is not. A platform tutorial. A news site. A vendor endorsement. A place to find windfall-style prompt recipes.
- What you should expect. New working notes weekly. Long-form explainers monthly. Glossary updates as needed. Dispatches on Saturday mornings with the week's operator-decision pattern essay, recap, and observation.
- Who it's for. Anyone running an agent for longer than a single conversation.