Principle 10delegation

Optimise for steering, not only initiating

The system should support users not only in starting tasks, but also in guiding, refining, reprioritising, and correcting work while it is underway.

Key Facts

Cluster: delegation
Primary risk: The only available control becomes repetition or restart, which is inefficient and often operationally unsafe.
Related examples: 6 library examples
Doctrine library: Internal doctrine reference

Why does this principle matter?

Prompting is an initiation mechanism. It is not, by itself, a sufficient control model for complex or consequential work. Users require the ability to steer ongoing activity without restarting the entire process.

Support intervention mid-process.

Allow constraints and priorities to be updated dynamically.

Preserve continuity when the direction of work changes.

Treat the user as an active governor of execution, not only as an originator of requests.

What failure does this principle prevent?

The only available control becomes repetition or restart, which is inefficient and often operationally unsafe.

AI as interface embellishment: A conventional product is given a text input and labelled intelligent, without any meaningful change in operational model.

Simulated autonomy: The system appears autonomous in language or presentation but cannot act with meaningful independence.

Opaque execution: Work occurs in the background without adequate status, accountability, or recoverability.

How do I apply this in practice?

The implementation library contains inspectable, runnable examples that show this principle applied in real agentic product code. Study the source paths, library choices, and runtime assumptions — not just the pattern name.

Implementation example

Level 1: Augmented LLM — Single API Call

One model call with structured output, system prompt, and context. No loops, no tools. (script · advanced)

Inspect implementation

Implementation example

Level 2: Prompt Chains & Routing — Deterministic DAGs

Multiple LLM calls in a fixed sequence. Code controls the flow, not the model. (script · advanced)

Inspect implementation

Implementation example

Level 3: Tool-Calling Agent — Scoped Autonomy

The agent decides which tools to call and in what order, but only within a fixed set of well-defined capabilities. (script · advanced)

Inspect implementation

Key Facts

Related examples