Running an Event Team With Agents During Kerala Festival Season

Kerala festival season does not arrive gradually. Temple festival dates land, Onam programmes go up, pooram schedules firm, boat race dates shift with the water, and a small team that was comfortable last month is suddenly behind on three things at once.

The short version of what we learned running EventifyPlus through that pressure: agents were genuinely useful for the repetitive shaping of information, and close to useless for the judgement calls that decide whether a listing is right. The teams that get value from agents in this kind of season are the ones that separate those two categories early.

Three things spike at the same time

The difficulty is not volume by itself. It is that three different kinds of work peak together.

Listings arrive faster, and they arrive messy. Organisers who otherwise depend on word of mouth send what they have, which might be a poster image, a WhatsApp forward, or a phone call with a date and a place name.

Organiser messages spike too, and they need answers from someone who knows the local context. A question about whether a Theyyam performance runs through the night is not answerable from a template.

Schedule changes are the worst of the three. A time moves, a venue shifts, rain changes a boat race, and every change has to reach people who already planned around the old information.

Where agent help held up

The reliable wins were all in shaping information that a human had already decided about.

Turning a loose organiser message into a structured draft listing worked well. Someone still checked it, but the gap between a message arriving and a reviewable draft existing shrank a lot, and that gap was where things used to get lost.

Consistency work held up too. Same festival, described five different ways by five organisers, normalised into language a reader can compare. This is dull work that people do badly when tired, which makes it a good fit.

Catching missing fields worked. An agent asking "this listing has no end time and no contact number" before a human looks at it removes a whole class of back and forth.

Where it did not

Anything that needed local knowledge failed in ways that were not obvious at a glance, which is the dangerous kind of failure.

Place names in Kerala repeat. Temple names repeat more. An agent working from text alone will confidently attach a festival to the wrong location, and the output reads perfectly fine to anyone who is not from that area. Only a person who knows the district catches it, and that knowledge does not travel: someone at home among the forts and tea hills of Palakkad is no help on a Kannur listing.

Timing conventions caused trouble too. An all night performance, a procession that starts before dawn, a programme that runs across two calendar days: these do not fit tidily into a start time and an end time, and forcing them into that shape produces listings that are technically populated and practically wrong.

Cultural weight is the third gap. Knowing which of six events on the same day is the one people will travel for is not in the text of the submissions. It is in knowing the place, the way you only learn what a town like Kollam offers beyond the backwaters by spending time there.

The rule that emerged

Agents draft, people decide. A human approves before anything reaches the public listing. It is the same instinct behind tools that advise but never act.

That sounds like a limitation, and during the busiest weeks it was the thing that kept quality from sliding. The approval step is fast when the draft is good, and the draft was usually good, so the throughput gain survived even with a person in the loop on every published listing.

The corollary is that the review screen matters more than the model. Show the person what changed, put the field most likely to be wrong at the top, and make it easy to send something back.

What we would set up differently next time

Three things, in order of how much they would have helped.

• Flag anything with an ambiguous place name before a human sees it. Do not let the agent pick. Let it say it cannot pick.
• Treat schedule changes as their own workflow. A change to a live listing is a different risk from a new listing, and it deserves a different review path.
• Keep organiser conversations attached to the listing record. When context lives in one person's phone, the season decides who the bottleneck is.

The part nobody plans for

Peak season ends, and the habits formed under pressure stay. Shortcuts taken in week three because everyone was tired become the way the team works in the quiet months.

Worth doing after any spike: read back a sample of what went out during the busiest fortnight, not to blame anyone, but to find where the review step got thin. That is where next season's problems are already sitting.