Skip to main content

Action Library and Controlled Context

At this stage of the system, execution authority has already been clearly defined.

The remaining challenge is not power, but context.

How does an AI understand the system
without being given unrestricted access to it?

BookiAI answers this with a deliberate abstraction: the Action Library.

The problem with raw context​

Many AI systems expose context by:

  • Dumping database records
  • Injecting raw JSON
  • Expanding prompts indefinitely

This creates several problems:

  • Context becomes inconsistent
  • Prompts become unmanageable
  • AI behavior becomes unpredictable
  • Security and privacy boundaries blur

More context does not mean better understanding.

Unstructured context is liability, not capability.

What an Action Library is​

An Action Library is a curated collection of system-exposed capabilities that AI can invoke.

Each action:

  • Has a clear name and purpose
  • Accepts structured inputs
  • Returns structured outputs
  • Enforces strict scope and permissions

Crucially:

Actions expose intent, not implementation.

AI never queries the database directly. It asks the system to perform an action.

Actions are read-only by default​

In BookiAI, the Action Library starts with read-only actions.

Examples include:

  • Fetching a chart of accounts snapshot
  • Retrieving recent ledger summaries
  • Resolving currency or tax context
  • Checking policy constraints

These actions:

  • Provide necessary context
  • Produce no side effects
  • Are safe to call repeatedly

This ensures AI can reason without modifying system state.

Controlled context beats full access​

By mediating access through actions, the system gains control over:

  • What information is visible
  • How fresh the data is
  • How much detail is exposed
  • Which rules are enforced

The AI sees the system through a governed lens, not as an open surface.

This prevents:

  • Data leakage
  • Hallucinated assumptions
  • Context drift across prompts

Action logs as observability​

Every action invocation is logged:

  • Which action was called
  • With what parameters
  • At what time
  • By which role (Generate or Review)

These logs enable:

  • Debugging AI behavior
  • Replaying decisions
  • Auditing context usage
  • Improving future prompts

Context usage becomes observable, not implicit.

Actions are not tools for execution​

An important boundary must be respected:

Action Library β‰  execution engine

Actions:

  • Provide information
  • Validate constraints
  • Explain system state

They do not:

  • Write ledgers
  • Change balances
  • Execute decisions

Execution remains exclusively under the Controller’s authority.

Why Action Libraries scale​

Because actions are explicit and versioned:

  • Prompts remain stable
  • Context evolves safely
  • New capabilities can be added incrementally

This prevents prompt explosion and enables long-term maintenance.

The Action Library becomes the system’s contract with AI.

What comes next​

With roles, contracts, assurance, controllers, and context in place, one final question remains:

How do we manage prompts and instructions
as long-lived, governed assets?

πŸ‘‰ Prompt Governance and Evolution

Written by ChatGPT, reviewed by the BookiAI team.