AGENT MANIFESTO

Agent Manifesto

What Is the Agent Manifesto?

The Agent Manifesto is a set of structured prompt files that bring order to AI instruction systems. It centralizes core rules, separates execution from orchestration, and keeps instruction files minimal and logical — scaling from small prototypes to large codebases.

Core Principles

Load Only What You Need

Default context should contain only what the current work needs to begin correctly.

Prefer

  • Small always-available instructions
  • Capabilities loaded when relevant
  • Clear entry points into deeper guidance

Avoid

  • Carrying task-specific guidance everywhere
  • Adding context as a substitute for design
  • Treating more instruction as automatically safer

Earn Complexity

The smallest structure that solves the real current problem wins.

Prefer

  • Direct handling for simple work
  • Generalization after the pattern is proven
  • Changes traceable to an actual requirement

Avoid

  • Speculative configurability and single-use abstractions
  • Starting with the heaviest structure available
  • Treating complexity as evidence of quality

One Artifact, One Job

Every artifact should have one clear responsibility it can be judged against.

Prefer

  • Artifacts describable in one sentence
  • Names that reveal responsibility
  • Explicit handoffs between decision and execution

Avoid

  • Mixed policy, procedure, and reference content
  • Work units that decide their own routing
  • Broad catch-all files

One Owner Per Concern

Every rule, constraint, and definition has exactly one authoritative home.

Prefer

  • A single owner for each rule, referenced from elsewhere
  • Auditing existing sources before adding another
  • Extending the rightful owner when it still fits

Avoid

  • Local copies of shared conventions
  • Parallel or competing authorities for the same concern
  • Creating a cleaner-looking duplicate

Make Behavior Explicit

Assumptions, success criteria, uncertainty, and stopping conditions should be stated rather than inferred.

Prefer

  • Declared assumptions and visible success criteria
  • Ambiguity surfaced before action
  • Blocking conditions with clear next steps when they fire

Avoid

  • Soft language for hard requirements
  • Rules that only advise where they should gate
  • Critical checks that can be skipped silently

Ask Before You Cut

Changes that can lose work, disrupt users, or reshape authority require explicit consent.

Prefer

  • Naming the risk before acting
  • Explaining the intended safe outcome
  • Treating consent as a requirement, scoped to what was approved

Avoid

  • Proceeding because a change seems obvious
  • Hiding risk inside a larger change
  • Treating approval for one change as approval for related changes