Open main menu
ToolsBlog
Say hi

The Building Blocks of AI Agent Systems

I simplified my stack into clear building blocks so commands, skills, subagents, question gates, and review inboxes work as one system instead of disconnected prompts.

Build
•
The Building Blocks of AI Agent Systems
Do not index
I have been iterating on both my personal operator setup and how PailFlow runs delivery work with coding agents. The big shift was simplifying the system into building blocks with clear jobs, so I am not rebuilding logic every time I switch context.

The architecture I run now

notion image

Entrypoints

Every run starts from one of three places: commands, chat, or scheduled jobs.
  • Commands trigger explicit workflows.
  • Chat handles interactive work.
  • Scheduled jobs run deterministic background loops.

Skills as the base unit

I treat skills as the operational logic unit. Instead of relying on one-off prompts or brittle command chains, I package repeatable behavior into skills that can be invoked from any entrypoint.

Workflow skills and simple skills

I split the layer into workflow skills and simple skills.
  • Workflow skills orchestrate multi-step execution.
  • Simple skills handle one focused capability.
Workflow skills can call multiple simple skills and branch into subagents when I need parallel depth.

Subagents for parallel depth

When I need speed and depth together, workflow skills delegate to subagents. I use this for parallel research and parallel task execution, then pull results back into one decision stream.

Question tool for human input

When a run needs approval or missing context, it should pause instead of guessing. That branch goes to the question tool so a human can provide input and keep the system grounded.

Review inbox for asynchronous oversight

If a run can produce an artifact without blocking, it lands in a review inbox. That is where I review in beats during the day: approve, reject, or trigger next step.

Queue and control plane

I keep my queue in Obsidian right now. The specific tool can change, but one visible queue is what keeps execution aligned to real priorities.

Why this model works for me

It separates reusable logic, orchestration, recurrence, and review.
That separation is what makes the system usable across both my personal workflows and PailFlow delivery operations.
I'm building PailFlow in the open and sharing how I use AI systems to scale a one-woman business.
If you work in client services and want to see how AI can increase your project delivery capacity, book a PailFlow Delivery Audit.

Written by

Lola
Lola

Lola is the founder of Lunch Pail Labs. She enjoys discussing product, app marketplaces, and running a business. Feel free to connect with her on Twitter or LinkedIn.