Path to agentization - MAIster

MAIster separates agentization into four levels of control. Each level adds a reusable system object and moves human attention farther from individual actions while preserving accountability.

The four levels

Level	MAIster object	Human role	System responsibility
Co-work	Scratch Run	Frame the problem, steer the session, decide what to keep	Isolate the workspace and record the runtime, conversation, files, cost, and outcome
Human in the loop	Flow	Approve permissions, answer forms, review plans and diffs	Execute a versioned graph and stop at declared human gates
Adversarial / human on the loop	Policy	Set risk boundaries and handle escalations	Apply checks, independent AI judgment, consensus, bounded rework, budgets, and promotion rules
Full agentic	Platform Agent	Define mandate, grants, budget, triggers, and review expectations	Run recurring or event-driven work with identity, Project Brain, memory, and a visible Run ledger

The practical path is therefore:

Scratch Run  →  Flow  →  Policy  →  Platform Agent
co-work         HITL     human on loop   full agentic

When to move forward

Promote work to the next level only when the previous level is predictable:

A Scratch Run reveals a repeatable procedure.
A Flow captures that procedure, its evidence, and its human decisions.
A Policy makes retries, checks, budgets, escalation, and promotion explicit.
A Platform Agent receives a bounded recurring mandate and launches the same governed execution objects instead of bypassing them.

The control plane keeps platform-agent Runs, child Runs, evidence, budgets, human interventions, and final outcomes addressable during full agentic work.