At 9:14 PM on a Tuesday, a marketplace panel quietly changes a button label. Every seller running browser automation against that panel now has a broken pipeline — and most will discover it tomorrow morning, after the SLA clock has run all night. The sellers running a rule engine fare no better: the rule fires, fails, logs an error nobody reads, and stops. What happens next is the cleanest definition of the difference between automation rules and an agentic OMS that I know.
“Automation” has become a word every tool claims and no tool defines. So this post is the definitional one: what rules actually are, what an agentic OMS actually is, the three tests that separate the two, and — because the word “autonomous” gets abused — what honest autonomy looks like, including the part where a human still clicks approve sometimes.
Rules execute steps. Agents own outcomes.
A rule is an if-then statement: if a new order lands, generate the label. If the label is generated, generate the slip. Chain enough of these together and you get workflow automation — genuinely useful, and a big step up from clicking through panels by hand. But the chain has a defining property: it knows its steps and nothing else. It does not know what the steps are for. If reality matches the script, the chain runs. The moment reality deviates — a panel timeout, a changed response, an order in a state the rule never anticipated — the chain either halts or, worse, skips silently and reports green.
An agentic OMS inverts the contract. You give it an outcome — all of today's AJIO and Meesho orders processed, documented, and manifested before cutoff — and the system decides the steps, monitors its own progress, and works through obstacles to deliver the outcome. When it cannot, it tells you precisely what it could not do and why. The unit of delegation changes from the step to the job. That single inversion drives every practical difference in the table below.
It is worth saying clearly that rules are not the villain of this story. Rule-based automation took sellers off the click treadmill, and for stable, narrow tasks — a stock alert at a threshold, a daily export at six — rules remain the right tool. The trouble starts when sellers stack dozens of rules into a load-bearing pipeline and then discover they have built something with the fragility of software and none of its guarantees. The maintenance burden lands on the one person least available to carry it: the founder, at night, after a panel update. The three tests below are how you find out, before committing your operations, which side of the line a tool actually lives on.
Test one: what happens when order 17 fails?
Here is the test I ask sellers to put to any tool claiming automation. You have 30 orders in today's run. Order 17 fails at label generation — the panel hiccuped, the courier allocation timed out, whatever. What happens to orders 18 through 30?
In a rule chain, the honest answer is usually one of two bad ones: the chain halts and 13 orders sit stranded behind the failure, or the chain skips order 17 silently and you learn about it from an SLA penalty two weeks later. In an agentic run, the failure is isolated: the other 29 orders complete, order 17 gets retried — possibly down a different path — and if it still cannot be processed, it gets escalated to you as a decision request with context: what failed, why, and the one choice needed from you. The difference in outcome is not subtle. It is 29 parcels on today's truck versus zero parcels past the failure point.
Test two: does it survive an interruption?
Marketplace panels are not stable substrates. Sessions expire mid-batch. Pages take ninety seconds on sale days. The network drops. A stateless rule chain, interrupted at order 14 of 30, restarts from zero — or worse, re-processes the first 13 and generates duplicate documents. Ask any seller who has manifested the same order twice how enjoyable the cleanup is.
Session persistence is the unglamorous engineering that separates a demo from a production system: the agent knows which orders are done, which are in flight, and which remain. Kill it mid-run and it resumes at order 14 — re-authenticates, re-verifies the state of the in-flight order against the panel, and continues. You will never see this feature on a pricing page, and it is half of what “autonomous” actually means. This is also why agentic order processing can run on a schedule at 7 AM without a human watching: surviving the substrate is the job.
The interruption test matters most precisely when volume is highest. Sale days triple traffic on every marketplace system at once: panels slow, sessions drop, timeouts multiply — the substrate is at its least reliable on the day your order count is at its most valuable. A pipeline that only works when the panel behaves is a pipeline that works on the days you least need it. Ask the vendor the blunt version of the question: if the connection dies at order 14 of 30, what exactly happens at order 14, and what happens to 15 through 30?
Test three: does it report outcomes or steps?
A rule engine's report is a log: step 412 succeeded, step 413 succeeded, step 414 failed. Technically complete, operationally useless — you become the integrator who reads logs and reconstructs what actually happened to your business. An agent reports at the level you delegate: 34 orders processed across AJIO and Meesho, documents generated, two manifests closed, one order needs a decision — stock mismatch on SKU KRT-204, here are your options. One is telemetry; the other is an employee's end-of-day summary. The report format sounds cosmetic until you realise it determines what your mornings look like.
Outcome reporting also changes what improvement looks like. Step logs accumulate; nobody mines them. An outcome report that says “order needed manual handling: stock mismatch” three times in a fortnight is pointing at a process fix — and a system that owns outcomes has every incentive to surface that pattern, because the exceptions are its workload too. The report becomes a weekly conversation about what to fix upstream rather than a forensic record you consult after something burns.
The autonomy spectrum — and where the human stays
Autonomy is a spectrum, not a switch. Most sellers travel it in stages: fully manual panel work; assisted ops with bulk tools; rule-based triggers; supervised agentic runs; and finally agentic operations, where the system owns outcomes across processing, documents, returns, and claims. Two honest observations about the far end of that spectrum:
- The routine loop genuinely needs no human. Accepting orders, generating labels and slips, closing manifests, tracking returns — these are deterministic judgment-free jobs, and a system that cannot run them unattended is not agentic. Robnu runs this loop for AJIO and Meesho today, on schedule.
- Rare, irreversible calls still route to you. Fully-autonomous claim filing is rolling out at Robnu now — and even there, certain submissions arrive as a one-click approval rather than firing silently. Filing a claim makes assertions in your name; the agent prepares everything and the human spends one click, not one evening.
- Distrust 100% claims. A vendor claiming zero human involvement across all operations is either not handling the judgment calls or not telling you who eats the mistakes. Honest autonomy is “the system does the work; you keep the decisions that are genuinely yours.”
Where a given seller should sit on the spectrum is a judgment about their business, not about technology. A seller with one channel, five SKUs, and a tight routine loses little at stage three. A seller running AJIO and Meesho together, with returns arriving daily and settlements landing on two calendars, is paying a real tax for every stage below four — paid in evenings, missed claim windows, and the occasional silent failure discovered too late. The spectrum is not a leaderboard. It is a map for deciding what your operation should stop asking of you.
What the agentic OMS difference buys a two-person team
The architecture argument matters because of what it does to a founder's week. Rules reduce clicking; they do not reduce responsibility. With a rule stack you remain the system's exception handler and integrator: you watch for silent failures, you restart broken chains, you read logs, you are on call for your own automation. The hours saved on clicking partially return as hours spent supervising the machinery — and the anxiety never leaves, because you know the rules do not know what they are doing.
Delegating outcomes changes the shape of the job. The morning ritual becomes a morning report. The 9 PM “did everything ship?” check becomes an evening summary that answers the question before you ask it. The skill a founder needs shifts from operating panels to reviewing work — the same shift that happens when you hire your first ops person, except this hire works every day at 7 AM, does not fall sick during sale week, and costs nothing right now. For a 2-person brand doing 5–25 orders a day across AJIO and Meesho, that is the difference between operations being the business and operations being a 15-minute agenda item in it.
There is a second-order effect worth naming: consistency compounds into protection. An agent that runs the loop every day also builds the evidence trail every day — documents tied to orders, timestamps on every step, returns tagged on arrival. When a deduction or a disputed return shows up weeks later, the case is already assembled. Rule stacks rarely deliver this because the evidence lives wherever each rule happened to leave it. Outcome ownership turns out to be an accounting position as much as an engineering one.
Where to start
If you are running rules today — or running panels by hand — the migration is not a leap of faith, because you can supervise before you delegate. Connect a channel, let the agent run the daily loop on schedule, and read the morning report against your own knowledge of the day for a week or two. Confidence in an agent is earned the same way it is with a new hire: by checking their work until you stop needing to. The full definition and architecture live on the agentic OMS page, and the how-it-works walkthrough shows the schedule-and-report loop end to end.
The economics make the experiment free in the literal sense: Robnu is free for everyone right now — every feature, every order, no card, no trial timer — and sellers under 25 orders/day stay free forever when paid pricing eventually launches. The question worth answering is not whether you can afford an agentic OMS. It is how many evenings of panel work you are still doing that a system should be doing for you.