The Strangest Data Shortage of 2026: Recordings of Software Being Used

The best-funded modeling problem in the industry — agents that operate software — is bottlenecked on a data type that barely existed as a category two years ago: trajectories of real computer use. Click-by-click, screenshot-by-screenshot records of tasks being completed in actual applications. This is a market briefing on agent trajectory data: why demand exploded, what’s actually available, and the four questions to ask any vendor claiming to sell it.

Why This Category Exists at All

Two shifts collided. First, computer-use agents went from research demo to product roadmap across the industry — an agent that books the flight, files the expense, reconciles the spreadsheet needs training examples of those tasks *as performed*, not as described. Second, the industry learned from text what happens when a capability outruns its supervision data: models trained on their own outputs drift, and the corrective is fresh demonstrations from real operators.

Hence the shortage. An LLM agent trajectory dataset — sequences of states, actions, and outcomes across real software — is expensive to produce (operator time, environment setup, failure-case capture), awkward to source from the open web (most screen activity is never public), and suddenly demanded by every fine-tuning team with an automation roadmap. Search interest in computer use agent training data has climbed steadily since the first computer-use demos shipped; the supply curve has not kept up.

What’s Actually on the Shelf

The category’s inventory, as it stands:

The licensed-corpus row is where Thordata’s datamall has quietly built something odd and valuable: an inventory that includes Claude conversation logs at 2,000-turn depth — extended, real multi-turn tool-using dialogue of exactly the shape agent fine-tuning consumes — alongside agent operation-log datasets covering task trajectories. For teams searching for an agentic workflow dataset for fine-tuning, the practical difference from the synthetic path is provenance: trajectories of real use carry the failure modes, recovery patterns, and long-horizon coherence that sandboxes systematically omit.

The Four Questions That Separate Trajectory Vendors

1. What’s in a step? A trajectory step should carry the observation (screenshot or state), the action (click, type, scroll, call), and the outcome. Vendors describing “task data” without that triple are selling something else.
2. Whose hands? Synthetic rollouts are fine and cheap — when labeled as such. “Operator-collected” claims should come with capture methodology. Mixing the two unlabeled poisons fine-tuning in ways that surface only in evaluation.
3. What’s the failure rate — on purpose? Trajectories of only successful tasks train agents that can’t recover. The good corpora include failed attempts and recovery paths *by design*; ask what fraction.
4. What does the license permit? This category’s records embed third-party software UI and user behavior. A per-record license scope that names commercial training, and a vendor who has thought about the display-of-UI question, is the difference between a dataset and a discovery request.

The Pattern From the Text Era, Ported Early

The teams fine-tuning agents now are reliving 2023’s text-supply scramble with one advantage: they can see it coming. The playbook that worked for text transfers directly — license breadth for base behavior, collect your own for your specific workflows, hold out fresh demonstrations for evaluation, and keep provenance attached so the enterprise procurement conversation doesn’t restart from zero. Every layer of that stack is rentable now; the agent era’s difference is that “collect your own” is the *default* rather than the fallback, because your workflows are the point.

FAQ

What’s a Claude conversation dataset actually good for?
Long-horizon, tool-using dialogue behavior: multi-turn coherence, tool-call timing, recovery from bad outputs. The 2,000-turn depth matters because most assistant conversation corpora are single-digit turns — they teach agents to start tasks, not to finish them.

Computer use agent training data — is this just screenshots?
Screenshots are the observation third of the triple. Without aligned actions and outcomes, screen recordings are video content; with them, they’re trajectories. The schema question in the four-question list exists precisely because “screenshots” gets marketed as “agent data.”

Synthetic vs. real trajectories — do we have to choose?
No, but label them. The strongest published results combine synthetic scale with real-behavior grounding — the failure mode isn’t using synthetic data, it’s not knowing which part of your corpus is synthetic.

We need trajectories for our internal app, which no vendor has.
That’s the collect-your-own layer: custom collection scoped to your workflows, riding the same infrastructure that serves the catalog side (residential collection across 190+ countries, published pricing from $0.65–$2.00/GB residential). And when your agents eventually face the outside world — searching, browsing, verifying — the SERP monitoring solution supplies the query-behavior layer agents need to reason over, at about $0.70 per 1,000 structured responses.

Where does search visibility fit an agent product?
Two places, both current blind spots for most agent teams: watching how users ask AI assistants about your agent category — the fastest-moving query vocabulary in software — and tracking where your agent’s own product surfaces in an answer-layer that increasingly recommends tools. Continuous SERP data crawling delivers both, and query-level monitoring across your category’s terms is the cheapest research feed in a category this new.

The briefing’s conclusion: agent data went from nobody’s product to everybody’s bottleneck in under two years, the shelf inventory finally exists — trajectories, conversation logs, operation records — and the teams that license it now are the ones whose agents will have seen a computer before their competitors’ do.