Trust and control
A worka is an assistant, not an autonomous decision-maker. It gathers, organizes, prepares, and recommends - and the platform is built so that humans own the decisions, approvals, payments, and accountability. Here is exactly how much rope you give it, and how you take it back.
The autonomy ladder
Autonomy on ezWorka is graduated, not all-or-nothing. These are the four modes defined in the platform contracts - you choose per worka and per capability, and you can move down the ladder at any time.
Every send, publish, payment, posting, or account change comes back to you as a pending approval with the full context. Nothing irreversible happens without your yes.
You keep: Every decision, every time.
You grant approval in advance for a specific worka, tool, or action class - with limits you choose, like a maximum number of runs or a spend cap per action. Anything outside the grant still comes back to you.
You keep: The boundaries: scope, caps, and expiry are yours to set and revoke.
A scoped grant lets a trusted worka execute inside a boundary you define - specific tools, accounts, or domains. Grants are recorded, time-boxed if you want, and revocable in one click.
You keep: The walls. A grant is never "everything"; it must name its scope.
Flip the emergency stop and workas stop executing - grants or not. Work in flight is blocked, and nothing runs again until you turn it back on.
You keep: The kill switch. It beats every grant on the ladder.
Guardrails
We label these honestly. 'Live today' is how hired workas behave right now; 'rolling out' is contract-defined behavior the platform is actively enforcing surface by surface.
Live workas are built on a prepare-then-ask convention: instead of executing a protected action (a send, an order, a payment instruction), the worka returns a pending approval for a human to decide. Ad campaign drafts, for example, are created paused.
A hire grants a named bundle of tools - nothing else. The license a worka runs under lists its entitlements explicitly, and a tool outside the list is simply not available to it.
The autonomy model is contract-defined: a grant must name its scope (worka, tools, accounts, or domains) and can carry execution counts, per-action spend caps, total caps, and expiry dates. Platform-wide enforcement of this policy layer is rolling out.
Every autonomy decision - approvals, rejections, grants, revocations, emergency stops, blocked executions - is designed to be recorded in an append-only audit log that no worka can edit or delete. The audit surface is part of the current platform build-out.
The plain-language promise
Where a worka is live today, protected actions do not execute silently: the worka prepares the work and returns a pending approval for a human to decide. Drafts stay drafts, campaigns are created paused, and payment instructions are prepared - never released - by a worka.
As the platform-wide policy layer rolls out, the same rule is enforced below the model for every tool a worka can reach: destructive and irreversible action classes default to requiring approval, unknown capabilities are refused, and only an explicit, scoped grant from you can change that.
If anything ever feels off, the emergency stop halts execution across your workas - grants or not - until you say otherwise.
Start in manual