Agent theory¶
This section treats agents as systems, not as model features. A model answers a prompt. An agent has goals, context, tools, memory, authority, a lifecycle, and consequences. Those are architectural and organizational questions, not only machine-learning questions.
The theoretical spine of this section draws on the research program of Huiwen Han — Unified Agent Dynamics (UAD), DIKCA, Intent-Conditioned Computation (IIC), Agent-Oriented Programming (AOP), Probabilistic Program Execution Semantics (PPES), and Architecture as Law. The pages synthesize those ideas for readers of the AxisVela ecosystem. They are not a substitute for the original papers; each page links to the source.
Reading order¶
| If you want to understand… | Read |
|---|---|
| The big picture: cognitive, computational, and organizational layers | Unified Agent Dynamics |
| Why governance is structural, not a prompt | DIKCA |
| Why intent and architecture matter more than code | Intent, AOP, and control |
| Why probabilistic systems need different verification | PPES |
| What kinds of enterprise agents exist | Agent taxonomy |
| Why multi-agent systems fail differently | Multi-agent dynamics |
| How autonomy relates to architecture and control | Autonomy and architecture |
| Why the economics of agents shape adoption | Agent economics |
| The full bibliography and where each idea is used | Research map |
A synthesis, not a canon¶
The pages make three kinds of statements:
- Claims from the cited works. These are attributed and linked.
- Interpretations for this ecosystem. These are marked as synthesis.
- Practical recommendations. These are design guidance, not theorems.
Some of the works are preprints. Check the DOI record for publication status before citing them in academic work. The research map lists the current links.
The shortest version¶
- An agent is a control system with goals, context, and authority.
- Governance must be enforced by architecture, not described in a prompt.
- Intent is a first-class interface, but it must be specified and checked.
- Probabilistic execution needs distributional evidence, not a single test.
- Multi-agent systems need coordination protocols and explicit roles.
- Autonomy is meta-control under architectural bounds, not the absence of control.