One Technician, 15 Records a Day, 12,000 Objects: What Happens When Four Agents Catalog at Once

Eight hundred working days. That is the number that quietly kills most collections-online projects, and it is worth writing on a whiteboard in the collections office before anyone talks about technology.

Twelve thousand objects have been photographed. The images sit on the drive. The board has been told the collection is going online. And one Collections Data Technician, working steadily and well, finishes about fifteen public-ready records a day. Twelve thousand divided by fifteen is eight hundred days — a little over three years, assuming she never gets pulled onto an exhibition install and never leaves. She will get pulled onto an install. She will probably leave, because that is what happens to people paid $22 an hour to do one task for three years.

The number that makes a collections-online launch impossible

The cruel part is that no single record is hard. It is not scholarship. It is seven small acts of tidying repeated twelve thousand times, never difficult enough to be interesting and never simple enough to have been automatable in the 2024 sense.

Here is what one public-ready record actually takes, in the order she does it:

• Clean the title. "Untitled (2 women, beach?)" becomes something a visitor can read, without inventing a title the curator never approved.
• Normalize the date. "ca. 1910s?" from a 1961 accession card becomes a structured range the website can sort and filter.
• Match the maker to an authority record — Getty ULAN for people, AAT for object types — so that "Wm. Hendricks, photog." and "William Hendricks" stop being two different artists on your website.
• Write sixty words of public description. Not curatorial prose. Reading-level-appropriate, factual, no speculation.
• Write alt text for the image, which your accessibility consultant has been asking for since the last website redesign.
• Check rights status: public domain, museum-owned, artist's estate, or unclear. Unclear means it does not publish.
• Translate the description into Spanish, because your service area is 31% Spanish-speaking and your last IMLS-funded evaluation said so in bold type.

Seven steps, roughly twenty-five minutes a record, and the ten thousandth takes exactly as long as the first. Nothing gets faster past about week three.

Slow because it is serial, not because it is hard

This is the distinction worth being precise about, because it determines whether 2026 helps you or not. A serial backlog is one where the work is not difficult, there is simply an enormous amount of it and only one person is allowed to be doing it at a time. Cataloging a photographed collection is the purest example in the museum field. Record 4,001 does not depend on record 4,000. They are in a line for no reason except that there is one technician.

Museums have tried to break the line. Crowdsourced transcription helped with handwritten material. Volunteer cataloging days produce enthusiasm and inconsistent data. A second technician costs $52,000 a year fully loaded and halves the calendar at best — 800 days becomes 400, still past the end of the strategic plan the board approved.

flowchart TD
  A["Registrar selects 12,000 photographed objects in TMS"] --> B["Job split by collection area"]
  B --> C["Lane 1: photographs and negatives"]
  B --> D["Lane 2: textiles and costume"]
  B --> E["Lane 3: tools and agricultural"]
  B --> F["Lane 4: works on paper"]
  C --> G["Merged drafts queued for review"]
  D --> G
  E --> G
  F --> G
  G --> H["Collections technician approves or kicks back"]
  H --> I["Published to collections online"]

What Agent Teams changed

The development to pin this on is Agent Teams, which arrived as a research preview alongside Claude Opus 4.6, and the same idea appearing across the other 2026 releases: several agents split one large job, work at the same time, and hand back merged results. Not one assistant working faster. Four of them working on different parts of the same pile simultaneously, then combining what they produced.

For a software team that is an interesting capability. For a museum with a backlog it is the whole ballgame, because the museum's problem was never quality per record — a competent assistant could already draft a decent sixty-word description in 2025. The problem was that it drafted them one at a time while a human sat and waited, which meant the human was still the bottleneck and the calendar barely moved.

Split by collection area and four lanes run at once: photographs and negatives, textiles and costume, tools and agricultural implements, works on paper. Each lane reads the accession card, the existing tombstone record, the photograph, and your own style guide, and produces the seven-step draft. They do not publish anything. They fill a review queue.

A Tuesday in the collections office, four lanes at once

The technician's day changes shape entirely. She arrives to a queue of 480 drafted records from overnight — 120 per lane. Her job is no longer to type; it is to judge. She works down the queue in a review view: title, date, maker match, description, alt text, rights flag, Spanish text, all on one screen with the accession card image beside it.

Most records take her forty seconds. Approve. Approve. Approve. Then one stops her: the maker was matched to a William Hendricks in the authority file who was working in Ohio, and yours was a studio photographer in Fresno. She rejects it, notes the reason, and every subsequent record in that lane inherits the correction. That is the part that did not exist in 2024 — the correction propagates instead of being retyped ninety times.

Roughly one record in seven comes back to her as genuinely uncertain. Rights status is the biggest offender: an unsigned photograph from 1948 with no accession paperwork gets flagged "cannot determine" rather than guessed, and it goes into a pile for the registrar to decide. That flagging behavior is the single most important thing to test before you trust any of this. A system that guesses at rights status will publish something you have to take down, and taking something down after a photographer's grandchild sees it is a very bad afternoon.

Three years or one spring: the arithmetic

Assumptions: 12,000 records, one technician at $22/hour loaded to $29, 6.5 productive hours a day, 15 records/day by hand, and review of drafted records at 120/day — conservative once a style guide settles, but plan for 120.

The money is real but it is the second-best argument. The best argument is the calendar. A collections-online launch that lands in five months can be tied to an exhibition opening, written into a grant report while the grant is still open, and shown to the board by the same director who promised it. One that lands in three years belongs to somebody else's tenure.

What must not go into the queue

Attribution stays with the curator. If your record says "attributed to" or "circle of," no drafted record should quietly upgrade that to a firm attribution because the description reads better that way. Lock those fields.

Culturally sensitive material comes out of the batch before it starts. Anything under NAGPRA review, anything a tribe has asked not be published, anything with human remains or funerary context in the record — pull those accession numbers out of the selection at the start and handle them through consultation, at consultation's pace. This is not a technical control, it is a list your registrar builds by hand, and it should be built before anyone runs anything.

Label voice is a real limit too. Drafted descriptions come back competent and slightly flat. If your museum has spent years developing an interpretive voice — plain-language, first-person community history, whatever your education director fought for — expect to spend the first two weeks rewriting the style guide until the drafts sound like your museum rather than like a catalog. Budget that as project time, not as a surprise.

And donor and provenance notes stay out of the public field entirely. The number of accession records containing a sentence like "Mrs. Whitcomb was quite firm that her sister not be told" is higher than you think.

Start Monday with 200 records from your least sensitive collection area — postcards, tools, local newspapers. Run one lane, review all 200 yourself, and count how many you would have published unchanged. Above 80%, add lanes. Below 50%, your style guide is the problem, not the technology.

Frequently asked questions

Will this write into TMS or PastPerfect directly?

Do not let it, at least not at first. Have it produce a review file a person approves, then import approved records on a schedule. Writing straight into the collections system removes the one checkpoint that makes this safe, and that system is your record of authority for insurance, loans, and audit.

Our records are inconsistent going back to 1958. Does that break it?

Inconsistency is the normal case and it is handled better now than it was two years ago — messy, partial, and contradictory source material is exactly what improved. What breaks it is silent guessing. Test specifically for whether it flags an unreadable accession card as unreadable instead of inventing a plausible date.

Can volunteers do the review instead of paid staff?

For descriptions and alt text, often yes, with a trained volunteer and a clear checklist. For rights status, maker attribution, and anything touching sensitive material, no. Those decisions carry legal and relationship consequences and they belong to the registrar.

What happens to the technician's job?

In the museums I have visited that did something like this, the technician is not cut — she moves to the work that was always waiting: location audits, the inventory nobody has finished since 2011, the loan paperwork that keeps slipping. If your plan is to cut the position, say so at the start, because she will figure it out by week two.

A note on the visitor side. The moment a collection goes online the questions change: callers ask whether a specific object is on view, whether they can see something in storage, whether their grandfather's photograph is the one on your homepage. CallSphere builds AI voice and chat agents that handle that inbound flow on the phone and on your website, answer from your own approved information, and route real research requests to the registrar instead of letting them pile up in a shared inbox.