Map, Automate, Break, Fix: How We Automated Our Own News and Jobs Desk

People ask us for an automation success story. Here is one we can tell honestly, because it is ours, warts and all.

Every morning, two pages on this website need to be fresh: our AI news page, a curated set of the day's most useful AI and tech stories, and our AI jobs board, a few hundred open AI roles pulled from many sources. For a while, someone did this by hand. It took time every day, and on busy days it simply did not happen.

Step 1: We mapped it

We described the job exactly the way you would in the Workflow Mapper. Trigger: every morning. Steps: gather stories from a list of publishers, keep only the ones that are genuinely about AI, remove the same story reported by different outlets, pick the best dozen, find an image and summary for each, update the page, publish. For jobs: collect from a set of boards, filter, tag seniority and region, remove duplicates, publish.

Tagging the steps made the decision obvious. Nearly everything was "could run itself". Only two things needed judgement: what counts as an AI story, and what counts as a duplicate. We wrote those down as explicit rules before building anything. For example, a story qualifies on its headline, not its body, because a body-text rule let in an unrelated essay that mentioned AI once in passing.

Step 2: We automated it

An agent now runs at 6 AM Sri Lanka time. It gathers, filters, de-duplicates, builds both pages and publishes them. The first week it looked like magic. Nobody typed "update" again.

Then we noticed something uncomfortable. The pages were updating every day, but some days they were worse. Fewer jobs. No news at all. The automation was not failing loudly. It was failing politely.

Step 3: It broke in ways nobody predicted

When we traced it, there were four separate problems, and none of them were "the AI got it wrong":

• It started too early. The schedule fired the moment the computer woke, before the internet connection was fully up. The run found almost nothing and published that. One morning the jobs board dropped from over 300 roles to 112.
• One source changed its format. A single jobs board began returning data in a slightly different shape, which crashed that whole source.
• A partial run overwrote a good one. A bad morning's output replaced yesterday's better page.
• Failures were invisible. The logs were stored somewhere that got wiped, so by the time we looked, the evidence was gone.

Step 4: We added guardrails, one per failure

1. Wait for the real world. Before doing anything, the agent now checks that several of the sites it actually reads are reachable. Checking just one was not enough: one morning that single site answered while every news publisher was still unreachable.
2. Refuse to publish something worse. If a run produces less than 60% of what is already live, it does not publish. An old but full page beats a fresh broken one.
3. Survive one bad source. Each source is handled on its own, so one broken feed no longer takes down the rest.
4. Make failures loud. Logs now live somewhere permanent and name exactly which sources failed and why.

Since then the two pages have refreshed every morning without anyone thinking about them. That is what success looks like for an automation: boring.

What we tell clients now

• Mapping gets you to the first version. It showed us what to automate and which two judgements to write down as rules.
• Reliability is a second map. After you map how the work flows, map how it can fail: bad inputs, missing inputs, partial outputs. Each one needs a check.
• Silent failure is the real enemy. An automation that stops is annoying. One that keeps running and quietly gets worse is dangerous. This is the "errors that repeat silently" risk in the 3 As of AI.
• Guardrails are features, not admissions of failure. The Workflow Mapper puts a guardrail on every recommendation for exactly this reason.

If you want to feel these failures safely before you build for real, Level 8 of the Automation Lab, "The cleanup", hands you a colleague's broken workflow and makes you fix it.

Learn it with Cocoon

This is the kind of lesson you only learn by running automations in a real business, and it is exactly what our trainers bring to every session: the build, the break and the fix. Cocoon trains teams across the region to map, automate and keep automations reliable, with trainers who do this work themselves. Meet the trainers or book a call.

Frequently asked questions