The State of AI Agent Use Cases · Q3 2026
A read of every use case in this directory: 583 concrete, task-sized things people point an AI agent at. No forecasts, no vendor decks. Just what the 583 recipes say about how these agents are triggered, how much they hand off, which tools show up, and what their authors warn will break.
71.4% of the 583 use cases run without being asked: 416 start on a schedule or fire on an event, and only 167 wait for you to click.
Each is a single, checkable statement drawn from the 583 use cases. Numbers link to a shareable page.
71.4% of the 583 use cases run without being asked: 416 start on a schedule or fire on an event, and only 167 wait for you to click.
Finding 1, share thisThe most common tool in the directory is OpenAI, wired into 181 of the 583 use cases (31%), just ahead of Google Sheets at 173.
Finding 2, share this88.5% of the use cases can be built in an afternoon or less, and only 11.5% count as a bigger project: most agent work is small.
Finding 3, share this16.8% of the 583 use cases keep a human in the loop, and only 485 run fully hands-off.
Finding 4, share thisThe failure authors warn about most is formatting, flagged in 191 of the 536 use cases that list a gotcha, ahead of auth and tokens at 147.
Finding 5, share thisEight views of the same 583 use cases. Every count comes straight from the recipe frontmatter, so it is reproducible from the dataset.
The 15 tools that appear in the most use cases. Names are normalized, so "Google Sheets" and "Sheets" count as one.
Whether a use case runs on a schedule, fires on an event, or waits for you to ask.
Whether the agent keeps you in control, drafts for your review, or runs fully on its own. In the current corpus this is uniform: every vetted use case is assistant-level, so the bar sits entirely on "you stay in control".
How much work it takes to stand each use case up.
The 12 largest categories of the 14 in the directory.
The tool pairs that share the most use cases: what tends to be wired up side by side.
Themes clustered from the "watch out" notes on the 536 use cases that list one. A note can hit more than one theme; 123 matched none of the eight.
How many distinct tools a single use case wires together.
Based on 583 vetted use cases in the directory as of 2026-07-10. Each use case is a real, task-sized recipe collected from community automation templates, YouTube, and Reddit, then filtered and rewritten so a normal person can follow it. Every entry carries a last-updated date, and anything that reads like a demo that never ran outside a screen recording stays out. More on how vetting works is on the about page.
Tool names are normalized before counting, so aliases of the same tool collapse into one. Trigger, autonomy, and setup-time labels are structured fields on every recipe, not guessed from text. The gotcha themes are clustered from the free-text "watch out" notes by fixed keyword matching, which is blunt on purpose: a single note can land in more than one theme.
One honest caveat: the autonomy field is uniform in this corpus. Every vetted use case is assistant-level, so the "how hands-off" chart has no spread yet. That is a real property of what has been let into the directory, not a measurement artefact, and it is why finding four is phrased as zero fully-autonomous agents rather than a percentage.
This is the baseline edition. Growth numbers (which tools and categories are rising or falling) start next quarter, once there is a prior snapshot to compare against. The raw counts behind every chart are published as a machine-readable snapshot.
The point of the directory is not to read about agents, it is to pick one task and hand it over today.
Browse all 583 use cases