Agents do the work. People make the calls.
One rule runs through all of Marline: an agent stages the change, and a person approves anything that spends money, changes a live listing, or reaches a customer or supplier in a new way. Here is how that works.
Connect, draft, review, graduate
Connect
Marline connects to your storefront, marketplaces, ad accounts, inbox, and warehouse. Every connection starts read-only. Agents learn how your store runs before they can change anything, and each later permission is asked for by name.
Draft
Each agent works from skills: written procedures for one job, like fixing a suppressed listing or planning a reorder. It does the work and stages the change with the numbers behind it. Claude reviews anything that needs judgment before it reaches you. Nothing goes live yet.
Review
Proposals reach you one at a time, with the reasoning and what was checked. Approve, edit, or reject. An edit becomes a rule the agent follows next time. A rejection stops the agent from retrying and records why.
Graduate
Evals test each agent against real past decisions. When a task's work passes those tests and your reviews, you can let it run on its own. Low-risk tasks only, one task at a time, and you can take it back whenever you want.
Daily budget, Holiday gift sets
Proposed $35
- Proposal
- Lower the daily budget on “Holiday gift sets” from $80 to $35.
- Why
- The season ended Dec 26. Since then the campaign has spent $412 for $96 in sales.
- Reviewed by
- Claude, before it reached you.
- If approved
- Applied in Amazon Ads and written to the log with this reasoning.
An example proposal, the way one reaches you. Try it: approve, edit, or reject.
Who decides what
Orchestration routes each task by what is at stake. The heavier the line, the more is committed.
Volume
Fast models
Low-cost models tag orders, reconcile counts, sort the inbox, and draft listing variants. High volume, easy to check, cheap to redo.
Judgment
Claude
Pricing, budget shifts, supplier problems, anything ambiguous. Claude, by Anthropic, weighs the evidence and writes a recommendation with its reasoning.
Commitment
You
Spending money, changing live listings, and anything that reaches a customer or supplier in a new way. A person signs off, every time.
Try it: route a task
Drag a task into the router, or press its Route button. See where it lands, and why.
Example data. Simulated.
Who decides
In the product: Orchestration routes each task by what is at stake: fast models for volume, Claude for judgment, and a person for anything that commits money, a live listing, or a relationship.
One log, readable by a person
Every action is written down: what changed, who did it, and why. Agents, Claude, and you, in one place. When something looks wrong next month, the reason is already there.
| Time | Who | What | Why |
|---|---|---|---|
| 09:14 | Fast model | Triaged a Walmart suppression on MRL-TEE-BLK-M as high priority | The item can't be bought |
| 09:14 | Listings agent | Staged a new Walmart title, 47 characters | Over 75 characters; prohibited claim “best” |
| 09:15 | Claude | Reviewed the title: needs a person | Changes a live listing |
| 09:21 | You | Approved the new Walmart title | |
| 09:21 | Listings agent | Updated the Walmart title | Approved by you at 09:21 |
| 09:30 | Creative QA agent | Fixed alt text on MRL-MUG-12OZ | Graduated task; ran on its own |
| 09:42 | You | Edited a daily budget from $35 to $40 | Saved as a rule: seasonal campaigns drop to $40 |
| 09:43 | Ads agent | Lowered Holiday gift sets to $40 a day | Approved with your edit |
Autonomy is earned, one task at a time
An agent never starts with autonomy. A task can graduate only when its work has passed evals on real past decisions and a run of your reviews without edits.
Graduation is narrow: one task, with a written scope. Fixing spelling in alt text can graduate. Changing a price, a title on a top seller, or a budget never does.
You can take it back at any time. The task goes back to asking you first.
Fix spelling in image alt text
Example- Approved runs
- 30 of 30 needed
- Edits in the last 20
- 0
- Evals
- Passed on 120 past decisions
- Scope
- Spelling and size in alt text only. Never prices, claims, or titles.
Runs on its own. Every run is still written to the log.
Try the review step
A queue of proposals from different agents. Approve with A, edit with E, reject with R, and watch each call land in the log.
Example data. Simulated.
Approval inbox
In the product: Every proposal reaches you one at a time with its reasoning and numbers. Your edits are saved as rules the agent follows next time, and every decision lands in one log.
Nothing goes live until you approve it.
Marline is pre-launch. Join the waitlist and we'll write when access opens.