Thoughtware

Organising Intelligence

Developing the vocabulary, primitives and first principles for organising cognitive capabilities into software systems that can scale reliably.

Thoughtware is a category of software with cognitive capabilities, built with one or more agents, powered by LLMs.

Agent

Owns a goal. Works through several judgments and actions until the goal is reached or it must stop. An agent has one or more Cognitive Units.

Cognitive unit

Owns one judgment. The smallest named piece of thinking in the system.

Thoughtware

An integrated system of agents and reusable cognitive units

Agent

Inquiry Agent

  • Interpret request
  • Maintain dialogue

Agent

Planning Agent

  • Build plan
  • Optimize sequence

Agent

Action Agent

  • Execute action
  • Handle exception

Agent

Inquiry Agent

  • Interpret request
  • Maintain dialogue

Agent

Planning Agent

  • Build plan
  • Optimize sequence

Agent

Action Agent

  • Execute action
  • Handle exception

Shared skills and cognitive units

  • Understand intent
  • Assess context
  • Judge priority
  • Check policy
  • Recommend response
  • Evaluate confidence
  • Learn from outcome

Software executed. People decided.

For most of its history, software applied judgments that had already been made. A rule lived in a condition. A calculation lived in a function. Same input, same output. The architecture was the encoding of thinking that happened before the program ran.

Then it predicted, still inside exact applications.

Machine learning recognised patterns: spam, fraud, a next purchase. Those models were specialists, embedded in otherwise deterministic products. The interface remained the product. The important judgments still belonged to people and to code.

Language made thinking computationally accessible.

General intelligence required abstraction. Language became that abstraction, one of the primary ways humans think, not only how they communicate. LLMs did not make language precise. They made ambiguity usable. Classify, plan, compare, criticise: the same surface.

Language is not simply how humans communicate. It is one of the primary ways humans think.

Designing in the Intelligence Age · Ch. 1

Capability arrived. Architecture did not.

Models can interpret, compare, and plan. That is not the same as having a judgment you can point at, test, version, reuse, or hand to a colleague. Without that edge, one word absorbs everything that thinks. The list below is not a progression of products. It is five architectures wearing the same name.

Mechanism vocabulary is precise about parts and silent about systems.

Introduction to Thoughtware · Ch. 1

Package thinking the way we packaged computation.

Functions did not give software new capability. They gave it something to point at. A cognitive unit does the same for judgment: one named decision, declared inputs, structured outputs, a versioned implementation, an evaluation suite. An agent owns a goal. Code stays exact.

The model is a source of cognition. Thoughtware is the system that gives cognition a place to live.

Introduction to Thoughtware · Ch. 5

Then intelligence can be organised.

Once a judgment has an edge, you can test it, reuse it, cost it, hand it to a colleague, and improve it without rewriting the application. That is what scale and reliability actually require, not a better LLM alone, but a place for each decision to live.