Learn

Core concepts

SyftOS is a governed home for a digital workforce. A handful of ideas explain how it all fits together, and why you can hand real work to AI agents without handing over control.

Organisations & tenancy

When you sign up you create an organisation, your tenant. Every agent, run, approval, integration and audit record belongs to it and is invisible to every other organisation: isolation is enforced at the data layer, not just the UI. You invite colleagues into your organisation and give each a role (below). One person can belong to more than one organisation and switch between them.

Agents & versions

An agent is a digital worker with a job to do: a brief, a model to think with, and a set of tools it may use. Crucially, an agent's behaviour isn't loose. It lives on versions.

The pay-off: a worker's behaviour is always traceable to an exact, frozen configuration. If you ever ask "what was this agent told to do when it sent that email?", the answer is a specific, unaltered version.

A version needs an objective, instructions and at least one tool before it can be published. Larger frontier models (the most capable OpenAI and Anthropic models) need the Growth plan or above. The model picker labels them and names the tier that unlocks them, and publishing is blocked if your plan doesn't include the model you've chosen.

Departments & workflows

A department is the fastest way to start: a ready-built team of agents, plus any pre-wired workflows, that deploy together in one step (for example Inbound Sales or Research). Everything arrives as drafts in your workspace, so a deploy is always safe to explore; nothing is live until you review and publish it. The whole install is atomic. If it can't all fit your plan, nothing is half-created.

The departments available today:

A workflow chains steps into a repeatable process and can run on a trigger: on demand, on a schedule, or from an incoming webhook another system sends. Schedules fire against your organisation's timezone, with daylight saving handled for you, so "9am" means 9am where you are. See Building workflows for the how-to.

Runs & dry-runs

A run is one execution of an agent against some input. Runs execute in the background, never blocking a page, and you can watch progress live.

Before trusting a new worker, dry-run it: it executes against sample input with every side effect simulated, so you can read its reasoning and see what it would do without anything actually happening. When you're satisfied, run it for real.

The approval gate

This is the heart of SyftOS. The instant a run wants to do something with a side effect in the outside world (send an email, create a task, move a deal, raise an invoice) it pauses and raises an approval request in the Approval Centre. Nothing happens until a person with the right permission approves it.

Audit & trust

Every meaningful event is written to an append-only audit log: a version published, a run completed, an action approved or rejected, an integration connected. It is never edited and never deleted, so it stands as a faithful record of what happened, who decided, and when. You can review and export it.

Over time each worker builds a track record you can see: how much it's done, and how its actions were decided. That history is what lets you widen a worker's remit with confidence: everything it does is gated and recorded.

Roles & permissions

Members hold a role in your organisation. Rather than one long ladder, the roles split building from deciding — so the person who proposes work and the person who approves it can be different by design:

The right to decide approvals (Owner, Admin and Approver) is deliberately separable from the right to build agents — and no agent ever holds it. That separation is what keeps a human firmly in the loop.

Your data & security

Connecting a system never hands over your password. For one-click connectors you authorise SyftOS through the provider's own consent screen; for key-based ones you paste a key. Either way the resulting credential is encrypted at rest, scoped to your organisation, and never shown again once stored. It never appears in logs, run history or the audit trail.

Permissions are least-privilege: each connector requests only the access its tools actually need. And because every side-effecting action still passes the approval gate, connecting a system grants the ability to act through it — never unattended action.

Next steps