Cognitiveering

Company · Our principles

Our principles

Bridges, payment systems and aircraft are engineered to fail safely and leave evidence. The way decisions are made, by people and now by AI agents, deserves the same discipline.

Document
Principles
Version
1.0
Principles
10

Preamble

Cognitiveering (cognitive + engineering) is the practice of designing how thinking turns into action: the inputs, the reasoning, the permissions, the actions and the proof of what actually happened. It builds on the established field of cognitive systems engineering and applies it to organisations in which people and AI agents decide together.

  1. 01

    Reasoning is not authority.

    A good argument, from a person or a model, does not grant permission to act. Policy does.

  2. 02

    Proposal is separate from execution.

    Whoever proposes an action should not be the only one who approves it.

  3. 03

    Approve exactly what will happen.

    An approval binds to the precise action, not to a general intention. Any change requires a new approval.

  4. 04

    Continuity belongs to the system.

    Commitments, schedules and state live in authoritative records, not in memory or summaries.

  5. 05

    Evidence over claims.

    "Done" means verified. A success message is not an outcome.

  6. 06

    Trust is assigned per source.

    Instructions and facts carry their origin. Untrusted content never grants new capabilities.

  7. 07

    Bound the absurd.

    Define normal ranges in advance, and pause when an action leaves them.

  8. 08

    Measure verified outcomes, not activity.

    Busy is not productive. Count results that can be proven, and what they cost.

  9. 09

    Review and calibrate.

    Every decision is a prediction. Score it, learn from it, and adjust.

  10. 10

    Autonomy is earned, per task.

    Independence is granted one capability at a time, as evidence of reliability accumulates.

The model reasons. The runtime owns continuity. Deterministic policy controls permissions. Tools perform actions. Verification establishes what actually happened.

Why now

AI agents are moving from chat windows into operations: ordering stock, signing contracts, hiring people, moving money. Public experiments show that they can do real work, and that they fail in recognisably human ways: eager to please, forgetful, too easily persuaded. The answer is not to stop building agents. It is to engineer the system around them, and to hold our own decisions to the same standard.

What follows from these principles

  • Veritor, a trust layer that puts principles 01–07 into practice for AI agents (an open-source core is planned).
  • The Agent Failure Library, which turns publicly reported failures into design lessons.
  • The Field Guide, a catalogue of patterns for reliable reasoning.

The Cognitiveering Briefing

Receive future revisions of the Principles.

Periodic. Incident analyses, research notes and Veritor releases. Unsubscribe at any time.

We will use your email to send the Cognitiveering Briefing. Unsubscribe at any time. See our Privacy Notice.