Platform

The T0-T4 constraint model

Five levels classify how much autonomy an agent is granted. Every agent is assigned a level before it acts, and promotions are recorded as validated events.

T0

Audit Trail

Every action is logged. Foundational, always on.

T1

Crypto Identity

Every agent gets a non-human identity. Foundational, always on.

T2

Guardrailed

Operational policies constrain what the agent can do.

T3

Policy-Driven

Advanced governance: rate limits and decision routing.

T4

Autonomous

Full self-operation under declared constraints.

Read this before you rely on the tiers The tier is a declared posture and promotion record. The policy engine enforces by tool risk and data classification — it does not branch on the tier. A T2 agent calling a high-risk tool with no matching policy is denied; a T4 agent calling the same tool is also denied. What the tier changes today is the governance posture recorded against the agent and the promotions it can move through.

Promotions

Tier promotions are a first-class event. Each one is validated against the published OpenAPI schema, and the valid set is exactly T0 through T4. A promotion records who moved the agent and why, so the history of an agent’s autonomy is reconstructible.

Not a guarantee of bounded autonomy “T4 agents act autonomously” is a description of intent, not a claim that the engine constrains them differently from T2. We would rather say that plainly than let a tier diagram imply enforcement that does not exist.