How to Fix a Mistake an AI Agent Made

You fix a mistake an AI marketing agent made by marking up the exact problem on the draft where it happened and sending it back to the same agent, not by scrapping the output and rewriting the brief from scratch. In Orbitable, that loop runs through the Dock: you pin the specific line or claim that's wrong, the agent revises against your note using the same shared context it started with, and the review step itself never costs credits.

Why re-briefing from scratch is the wrong first move

When something in an agent's output is wrong, the instinct is to close the draft and write a fresh, more detailed prompt. That throws away the one thing you actually have: a specific, located error. A new brief asks the agent to guess again at everything, tone, structure, the facts you didn't flag, when only one sentence or one number needed fixing. It also means you're typing out context the agent already has, since every Orbitable agent works from a shared world model per customer, your ICP, brand voice, products, competitors, and uploaded knowledge base. Re-briefing ignores that context instead of correcting against it.

The mistake usually isn't "the agent doesn't understand your business." It's "this one draft has one wrong claim, one off-brand phrase, or one missed step." Treat it as that, not as a reason to start over.

The correction path we actually use: markup, resubmit, resolve

Because we run 50 specialist agents daily across 10 squads, we've had to build an actual loop for this rather than leave it as a support ticket. The Dock is the surface it runs on. A draft lands there, you or a client stakeholder sees it, pins a markup directly on the part that's wrong, and sends it back. The agent that produced the draft picks up the revision with your note attached and its original context intact, not a different agent starting cold.

Requests raised this way carry SLA targets, so a markup isn't a message that sits unanswered, it's a tracked item with an expected turnaround. Once you approve the revision, it moves into the Library rather than staying a loose file you have to hunt for later.

Does fixing a mistake cost credits?

Reviewing a draft, marking it up, and sending it back costs nothing. Credits meter the work agents do, not the act of you checking that work, so pinning an issue and resubmitting it is free every time. The agent's revised pass is new output, the same way the first draft was, so it draws credits the same way any agent task does. What you're not paying for is your own time spent reviewing, catching, and flagging, which is the part that usually gets expensive when it's a person doing the checking manually across a dozen tools.

This matters for a specific reason: if review cost credits, teams would under-review to save budget, and mistakes would ship. Making the review step free is what makes the markup loop something people actually use instead of avoid.

Who does the fixing: you, the agent, or the Dispatcher

The Dispatcher orchestrates the fleet, it routes work, sequences agents, and keeps missions moving, but it isn't a 51st specialist and it doesn't rewrite content itself. When a markup comes back, it goes to the specific specialist agent that produced the draft, because that agent already has the working context for that piece: what it was trying to do, what source material it used, what brief it followed. That's a different job from starting a brand-new task, and it's why sending the correction back to the same agent gets you a tighter fix than routing it as a fresh request.

When the mistake is bigger than one draft

Sometimes what looks like one bad output is actually a bad input. If an agent keeps getting a fact, a persona detail, or a tone wrong across multiple pieces, the fix isn't another markup, it's the world model. Every agent for a given customer draws from the same ICP, brand voice, product facts, competitor list, and knowledge base. If one of those is stale or wrong, every agent touching that topic will repeat the error, because they're all working from the same shared source.

This is the direct upside of the shared-context design: a correction made once at the world model level compounds forward the same way an improvement does. You don't have to remember to tell ten agents the same thing ten separate times.

What good markup actually looks like

A markup that says "fix this" gives the agent nothing to work against. A markup that names the specific problem, this number is wrong, this claim isn't in our knowledge base, this phrase isn't our voice, gives it something precise to revise. If the error is factual, point at the correct source or figure rather than describing the feeling that something's off. If it's a pattern you've seen before, note that too, since a recurring note is often the signal that the fix belongs in the world model rather than in one more markup.

FAQ

Does correcting an agent's mistake cost extra credits?

Reviewing and marking up a draft costs nothing, since credits meter the work agents produce, not the act of checking it. The agent's revised pass after your markup is new work and draws credits the same way the original draft did. What you save is the cost of re-briefing from zero and paying for a whole new attempt.

Who actually reviews the mistake, a person or another agent?

A person reviews it, you or a client with Dock access, by pinning the specific issue directly on the draft. Orbitable doesn't route a mistake to a second agent to fact-check the first; the correction loop runs through human markup and the original agent's revision.

Can a client fix a mistake without me being involved?

Yes, if you've given them a Dock seat, which is client-facing and review-only. They can pin markup directly on a draft and raise it as a request with an SLA target, and it routes back to the agent the same way your own markup would. Internal team notes stay hidden from that client view either way.

What if the same kind of mistake keeps happening across different pieces?

That's usually a sign the problem lives in the shared world model, not in any single draft, your ICP details, brand voice guide, or knowledge base entry might be stale or wrong. Fixing it there corrects the source every agent draws from, rather than patching the same error draft by draft.

Does fixing one agent's mistake help with any of the other 50 agents?

Only if the fix goes into the shared world model rather than staying as a one-off markup on a single draft. Since every agent works from the same ICP, brand voice, product facts, and knowledge base, a correction made at that level applies the next time any agent touches the same topic.