Where Humans Must Stay in the Loop With AI Agents
Two checkpoints decide where a human has to step in when you are running AI marketing agents: approving the plan before work starts, and approving the output before it ships. In Orbitable those checkpoints have names, Autopilot proposes the weekly plan and stops, and the Dock is where every draft gets marked up, sent back, or approved into a Library before it goes anywhere near a client or a live channel.
That is a deliberate design choice, not a caveat bolted on afterwards. If you are evaluating any AI marketing system, the question to ask is not "does it have human oversight" (everyone says yes), it is "name the exact moment the system stops and waits for a person." Below are the moments, why each one exists, and what is genuinely safe to leave running.
Why "human in the loop" needs a location, not a promise
Most AI tooling markets "human in the loop" as a philosophy. It rarely tells you where the loop actually closes. A vague promise of oversight is worth nothing if the system can draft, approve, and publish inside the same process with nobody outside it. The useful version of this question is mechanical: where in the pipeline does the software have to stop, and who has to say yes before it continues.
In Orbitable there are exactly two such stops. The first is before a week of work begins. The second is before any single piece of work leaves the building. Everything in between, the actual drafting, the research, the analysis done by the 50 specialist agents across the 10 squads, runs without a human clicking through each step, because reviewing intermediate work is not where the risk sits. The risk sits at the two decision points either side of it.
Checkpoint one: Autopilot proposes, it never approves
Autopilot looks at your world model, your ICP, brand voice, product facts, competitors, and whatever you have uploaded to the knowledge base, and proposes a weekly plan: one focus, 3 to 5 missions, each with a why-now, a carry-over list of what did not get finished last week, and an explicit NOT-doing list of what it is deliberately skipping.
It proposes. It does not start the missions itself. Approval happens in the Dock, which means a human reads the why-now for each mission, checks the NOT-doing list is defensible, and only then does the week's work actually begin. This matters most for the parts of a plan that are easy to get wrong silently: chasing last week's carry-over instead of a genuinely more urgent opportunity, or picking a focus that fits the data but not the moment your business is actually in. Autopilot is available on paid plans, not the free tier, because a proposal worth reviewing is a paid feature, not a gimmick.
Checkpoint two: the Dock is where work ships
The Dock is the second gate, and it is the one that actually stops something going out into the world. A client or internal stakeholder sees what is waiting for them, can pin comments directly on the draft, send it back for revision, or approve it into a Library. New requests raised in the Dock carry SLA targets, so the review loop has a clock on it rather than sitting indefinitely.
Two details make this a real gate rather than a formality. First, internal team notes are never visible to a client in the Dock, so the review surface a client sees is clean, not a leak of your working notes. Second, the Dock is included on the Agency and Enterprise plans and available as a $49/mo add-on on Founder and Team, which means it is a real feature with a real cost, not a checkbox that is always on by default regardless of what you are running.
What agents are allowed to do without a human watching
Not every step needs a person. The drafting, the research, the competitive analysis, the outreach sequencing done inside a mission, once a mission has been approved, run without a human approving each sub-step. The Dispatcher orchestrates the 50 specialist agents (it is the orchestrator, not a 51st specialist), and that orchestration happens because reviewing every intermediate output would defeat the point of having agents at all.
This is where credits come in as the honest signal of where work is actually happening. Credits meter the work agents do. Reviewing that work, in the Dock or anywhere else, never costs credits. That split is not accidental. It tells you the system's own economics agree with where the human checkpoint should sit: on the decision, not on the labour.
Where the shared world model raises the stakes of getting the gates right
Every agent in Orbitable shares one world model per customer, your ICP, brand voice, products, competitors, and uploaded knowledge base. That shared context is the actual point of running a fleet rather than a single tool, because research one agent does compounds into work every other agent produces later. It also means a bad assumption entering the world model does not stay contained to one agent's output, it propagates into the next mission, the next draft, the next Autopilot proposal.
That is exactly why the two checkpoints matter more, not less, as the fleet gets more capable. The Dock and Autopilot's proposal-only design are not there because the agents are unreliable. They are there because a shared context means an error compounds the same way a good insight does, and the two gates are the only points where a human can catch that before it spreads.
Comparison: what runs unattended vs what needs a human
FAQ
Does Autopilot ever publish or execute work without approval?
No. Autopilot only proposes a weekly plan, one focus with 3 to 5 missions, a why-now for each, a carry-over list, and a NOT-doing list. Approval happens in the Dock, and missions do not start until that approval is given.
What is the Dock and who can see what inside it?
The Dock is a client-facing review surface where a client or stakeholder sees what is waiting, marks up drafts with pins, sends work back, or approves it into a Library. Internal team notes are never visible to a client, so the review view stays clean. It is included on Agency and Enterprise plans and available as a $49/mo add-on on Founder and Team.
Do I have to review every step an agent takes?
No. Once a mission is approved, the specialist agents and the Dispatcher that orchestrates them carry out the research, drafting and analysis without a human approving each sub-step. The two points that require a human are approving the plan (Autopilot) and approving the finished work (the Dock).
Does reviewing work in the Dock cost credits?
No. Credits meter the work agents do to produce a draft or complete a mission, but reviewing, marking up, or approving that work in the Dock never costs credits.
Is the Dock available on every plan?
It is included on the Agency and Enterprise plans. On Founder and Team it is available as a $49/mo add-on. A Dock-only seat for a client does not require a full internal seat.