The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
Error code: DatasetGenerationCastError
Exception: DatasetGenerationCastError
Message: An error occurred while generating the dataset
All the data files must have the same columns, but at some point there are 2 new columns ({'agent_question', 'definition'}) and 5 missing columns ({'source_url', 'handling', 'checked', 'provider', 'quote'}).
This happened while the csv dataset builder was generating data using
hf://datasets/hal9/agent-cloud-terms/terms.csv (at revision a9af9f406ff1e422bc6810a7e626179ec2ad7e9a), ['hf://datasets/hal9/agent-cloud-terms@a9af9f406ff1e422bc6810a7e626179ec2ad7e9a/providers.csv', 'hf://datasets/hal9/agent-cloud-terms@a9af9f406ff1e422bc6810a7e626179ec2ad7e9a/terms.csv']
Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1856, in _prepare_split_single
writer.write_table(table)
~~~~~~~~~~~~~~~~~~^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 811, in write_table
self._write_table(pa_table, writer_batch_size=writer_batch_size)
~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 822, in _write_table
pa_table = table_cast(pa_table, self._schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
term: string
definition: string
agent_question: string
-- schema metadata --
pandas: '{"index_columns": [{"kind": "range", "name": null, "start": 0, "' + 633
to
{'term': Value('string'), 'provider': Value('string'), 'handling': Value('string'), 'quote': Value('string'), 'source_url': Value('string'), 'checked': Value('string')}
because column names don't match
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1379, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 956, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1702, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1858, in _prepare_split_single
raise DatasetGenerationCastError.from_cast_error(
...<4 lines>...
)
datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
All the data files must have the same columns, but at some point there are 2 new columns ({'agent_question', 'definition'}) and 5 missing columns ({'source_url', 'handling', 'checked', 'provider', 'quote'}).
This happened while the csv dataset builder was generating data using
hf://datasets/hal9/agent-cloud-terms/terms.csv (at revision a9af9f406ff1e422bc6810a7e626179ec2ad7e9a), ['hf://datasets/hal9/agent-cloud-terms@a9af9f406ff1e422bc6810a7e626179ec2ad7e9a/providers.csv', 'hf://datasets/hal9/agent-cloud-terms@a9af9f406ff1e422bc6810a7e626179ec2ad7e9a/terms.csv']
Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
term string | provider string | handling string | quote string | source_url string | checked string |
|---|---|---|---|---|---|
agent-native cloud | Cloudish | The example the term is defined by: an agent gets its own key with one call, and spend is prepaid and metered per second. | An AI agent creates its own API key with one call and no signup; credits are prepaid and metered per second, so spend never exceeds the key's balance. | https://cloudish.ai/skill.md | 2026-11-16 |
agent self-signup | Cloudish | An agent creates its own API key with one call and no signup or card; every other call authenticates with that key. | Creating a key itself needs no credential β every other call below authenticates with the key it returns. | https://cloudish.ai/skill.md | 2026-11-16 |
spend cap | Cloudish | Prepaid balance; the key never goes below zero, and its running apps stop within a minute when it runs out. | A key never goes below zero. Usage is charged only up to the balance the key holds, and whatever runs past it is never owed. | https://cloudish.ai/credits.md | 2026-11-16 |
spend cap | Railway | A hard limit takes all workloads offline when usage reaches it (a custom email alert only warns). | Once your resource usage hits the specified hard limit, all your workloads will be taken offline to prevent them from incurring further resource usage. | https://docs.railway.com/reference/usage-limits | 2026-11-16 |
spend cap | Vercel | Team and project budgets can pause production deployments at the budget; pausing is not instantaneous. | Team and project budgets can automatically pause production deployments when the budget is reached. Pausing is not instantaneous. | https://vercel.com/docs/spend-management | 2026-11-16 |
spend cap | Google Cloud Run | Spend cap budgets pause one eligible service in one project when enforced; overages are still billed. | When a spend cap budget is enforced, usage of your specified services is automatically paused for the remainder of the budget period. The enforcement of spend caps isn't instant and any cost overages are billed as normal. | https://docs.cloud.google.com/billing/docs/how-to/budgets-spend-caps | 2026-11-16 |
spend cap | AWS | Budgets alert and can apply an action, but the information is updated up to three times a day. | AWS Budgets information is updated up to three times a day. Updates typically occur 8β12 hours after the previous update. | https://docs.aws.amazon.com/cost-management/latest/userguide/budgets-managing-costs.html | 2026-11-16 |
spend cap | Azure | A spending limit exists, but not on pay-as-you-go or commitment-plan subscriptions. | The spending limit isn't available for subscriptions with commitment plans or with pay-as-you-go pricing. | https://learn.microsoft.com/en-us/azure/cost-management-billing/manage/spending-limit | 2026-11-16 |
prepaid compute | Cloudish | One credit is one US dollar; the balance cannot go below zero. | One credit is one US dollar. | https://cloudish.ai/instances.md | 2026-11-16 |
child key | Cloudish | Minting a key with a key makes it a child; fund it by transfer, and cap it with a receive limit. | Minting a key with a key makes it a child of that key β same account, no free grant of its own. Hand one to a sub-agent β¦ and fund it by transfer; a receive limit β¦ caps what it can get. | https://cloudish.ai/credits.md | 2026-11-16 |
claim link | Neon | An agent provisions a project with no account; a human claims it later, and unclaimed projects expire in 72 hours. | Unclaimed projects expire in 72 hours (project.expires_at) and are capped at 100 MB storage and 1 GB transfer. | https://neon.com/docs/reference/claimable-neon | 2026-11-16 |
claim link | Cloudish | A claim link attaches an agent-created key to a person's account; it works once and expires in 30 minutes. | The link works once and expires 30 minutes after it's minted. | https://cloudish.ai/skill.md | 2026-11-16 |
scale to zero | Cloudish | sleepAfterMinutes 1 to 1440, default 15; set it to never to disable sleeping. | sleepAfterMinutes (1 to 1440; 15 if you choose neither) puts it to sleep after that long without requests. | https://cloudish.ai/skill.md | 2026-11-16 |
scale to zero | Render | Free web services spin down after 15 minutes without inbound traffic. | Render spins down a Free web service that goes 15 minutes without receiving any inbound traffic. | https://render.com/docs/free | 2026-11-16 |
scale to zero | Railway | Serverless sleeps a service after at least 5 minutes without outbound traffic. | For Railway to put a service to sleep, a service must not send outbound traffic for at least 5 minutes. | https://docs.railway.com/reference/app-sleeping | 2026-11-16 |
scale to zero | Fly.io | Auto-stop can stop or suspend Machines when the app is idle for several minutes; resuming a suspended Machine takes a few hundred milliseconds. | auto_stop_machines: The action, if any, that Fly Proxy should take when the app is idle for several minutes. Resuming a suspended Machine takes a few hundred milliseconds. | https://docs.fly.io/launch/autostop-autostart | 2026-11-16 |
scale to zero | Google Cloud Run | Instances scale to zero when idle to reduce costs. | β¦handle traffic spikes and scale to zero when idle to reduce costs. | https://docs.cloud.google.com/run/docs/about-instance-autoscaling | 2026-11-16 |
scale to zero | Heroku | Eco web dynos sleep after 30 minutes without web traffic. | If an app has an Eco web dyno, and that dyno receives no web traffic in a 30-minute period, it sleeps. | https://devcenter.heroku.com/articles/eco-dyno-hours | 2026-11-16 |
persistent vs ephemeral storage | Cloudish | A volume of 1, 5 or 20 GB, attachable to one container only; kept (and not billed) even at zero balance. | volumeSizeGb must be one of 1, 5 or 20. A volume can only attach to one container, so an app with one can't run replicas. β¦ a zero balance stops it β¦ but the volume, with everything on it, is kept and isn't billed. | https://cloudish.ai/skill.md | 2026-11-16 |
persistent vs ephemeral storage | Render | The filesystem is ephemeral; local changes are lost on redeploy, restart or spin-down, and a persistent disk is paid-only. | Free web services have an ephemeral filesystem. This means that any changes to your web service's filesystem β¦ are lost every time the service redeploys, restarts, or spins down. You can preserve local filesystem changes by attaching a persistent disk, but Free web services cannot. | https://render.com/docs/disks | 2026-11-16 |
persistent vs ephemeral storage | Google Cloud Run | The filesystem is in-memory; data written to it does not persist when the instance stops. | It is an in-memory file system, so writing to it uses the instance's memory. Data written to the file system doesn't persist when the instance stops. | https://docs.cloud.google.com/run/docs/container-contract | 2026-11-16 |
agent-native cloud | null | null | null | null | null |
agent self-signup | null | null | null | null | null |
spend cap | null | null | null | null | null |
prepaid compute | null | null | null | null | null |
child key | null | null | null | null | null |
claim link | null | null | null | null | null |
scale to zero | null | null | null | null | null |
persistent vs ephemeral storage | null | null | null | null | null |
agent-cloud-terms
Eight terms for "clouds that AI agents use", each with a one-sentence definition and, per provider, what that provider's own docs say β quoted, sourced and dated. For agent builders, evaluator authors and writers who need a machine-readable, citable table.
Files
| File | What it is |
|---|---|
terms.csv |
The eight terms, one row each: term, definition (the one-sentence definition used across the other pieces), agent_question (what an agent asks to check it). |
providers.csv |
One row per term and provider: term, provider, handling (how the provider handles it, in plain words), quote (the exact sentence from the provider's docs), source_url, checked (date the quote was read). |
Quick start
from datasets import load_dataset
# One config per CSV file (named after the file)
terms = load_dataset("hal9/agent-cloud-terms", "terms", split="train")
providers = load_dataset("hal9/agent-cloud-terms", "providers", split="train")
Or read the CSVs directly from the repository.
The eight terms
- agent-native cloud β a cloud an AI agent can sign up for, pay for and deploy to with no person in the loop.
- agent self-signup β an agent creating its own credential with no human signup step.
- spend cap β a hard ceiling the work stops at, unlike a budget alert, which only warns.
- prepaid compute β compute paid up front from a balance that cannot go below zero.
- child key β a delegated credential under a parent key, funded by transfer and capped by a limit.
- claim link β a single-use link that attaches something an agent created to a person's account later.
- scale to zero β running no instances (and charging no compute) while idle, then starting one on the next request.
- persistent vs ephemeral storage β whether files survive a restart or redeploy.
Providers covered where their docs say something: Cloudish, Railway, Render, Fly.io, Vercel, Google Cloud Run, AWS, Heroku, Neon.
How it was built
On one date (see checked in each row), each provider's public documentation page named in source_url was read and the sentence that states the behaviour was copied verbatim into quote. The handling column restates it in plain words. "Agent-native cloud" is our own framing, not yet a standard term; the row for it is labelled accordingly.
Who made it
Written by Cloudish (cloudish.ai), by Hal9, which runs a cloud an AI agent deploys to. Cloudish chose the terms and appears as one provider among the others; rows where another provider does better are kept in, because the point is a fair table, not an ad.
Limits and biases
- Cloudish chose the terms. The eight words are the vocabulary Cloudish uses to describe its category, not an industry-standard list.
- Docs can differ from the product. Every row quotes documentation, checked on the date shown; a provider may have changed behaviour since.
- Provider coverage is uneven. Some terms have many provider rows (scale to zero), others only the providers that document them (child key, prepaid compute β mostly Cloudish, because few others document those exact ideas).
- "Agent-native cloud" is not vendor-neutral. It is explicitly marked as our framing.
Report a wrong row
Open a Discussion on this dataset page with the term, provider, and the corrected quote plus its URL and date. We update the row and bump the version tag.
License
The definitions, handling text and structure: CC BY 4.0. The quote column is a short quotation from each provider's own documentation for the purpose of comparison, with its source and date.
Citation
Cloudish (Hal9) (2026). agent-cloud-terms: eight agent-cloud terms with each provider's documented behaviour. Hugging Face dataset. Quotations checked 2026-11-16.
- Downloads last month
- 33