Our approach

What is “procedure”?

In any business, a workflow is a large task made of smaller tasks. The way these smaller tasks come together is usually called a “procedure” (how you “proceed” from one task to the next).

Procedures can be linear

(e.g. make a shot of espresso, put the espresso in the cup, pour some milk in the cup, stir)

or they can be nonlinear

(e.g. check if the invoice has been paid. If it has, enter the payment date and confirmation number into a spreadsheet. If it hasn’t, draft a follow-up email, then send that email)

Composite vs Agentic Systems

In the age of AI, workflow automations typically fall into one of two categories: “Composite” Systems and “Agentic” Systems.

The main difference:

In an agentic system, the LLM has authority over the procedure.

In a composite system, the procedure has authority over the LLM.

Composite systems were invented before agentic systems (2021 vs 2024). Now, in 2026, agentic systems are much more popular. We think this was a bad choice. Agents are broadly useful, but only the best choice in specific niche situations (e.g. personal assistants).

Over the past two years, the proliferation of agentic systems over composite systems for workflow automation has led to some major consequences:

  • High spend on API credits and tokens

  • Low visibility and monitorability

  • Procedural failures end up becoming high-level security breaches

  • Difficulty predicting if a system will work until it’s in production

  • There is no sense of control/responsibility over the agentic process: no single person “owns” the agent or the procedure, and oftentimes no single person fully understands how the whole process works (even the developers!)

We build composite AI systems to avoid these issues.

What is a composite AI system?

Two claims about workflows underlie the composite system design:

  1. A well-clarified procedure should usually be followed. Straying from this procedure should require escalation and approval.

  2. Different individual tasks require different types of solutions.

A composite system starts with a workflow, broken down into a collection of individual tasks and a well-clarified procedure. It implements this procedure as a deterministic program routing information between individual tasks.

Each individual task is assigned an appropriate solution. Some example solution types are:

  • An automated call to an LLM or other AI model

  • A classical machine learning model fit for that individual task

  • A regular deterministic computer program

  • A person completing the task or verifying an output

  • Escalation and alerts

Solutions are configured specifically for their individual tasks:

  • A classical ML model is trained on human-labeled results for a task

  • The size/cost of an LLM call is determined by the difficulty of its task

  • Deterministic programs are selected for deterministic tasks.

Choosing the right solution for the task is analogous to choosing the right personnel for one component of a workflow. Everything runs efficiently if the right people are doing the right tasks; the same is true for technical solutions.