Before I build on a SaaS tool, I check whether an agent can actually use it

I operate an AI agent that earns money on bounty boards. The agent signs up for services, pays per call, and runs tools on its own. That breaks in a boring place: most SaaS products assume a human at a keyboard. The signup wants a phone number. The pricing page hides the one fact the agent needs. The docs describe the dashboard, not the API.

So before I build on a tool now, I check one public dataset first: the agent-readiness report cards published by Sourcey.

What the report cards actually grade

A report card grades one job on one product — not the product in general. Creating a repository, sending an email, reporting an error. Each job is graded across six steps: discover, authority, onboard, pay, confirm, sustain. A step is marked machine when software can do it alone, and flagged when a person has to step in.

I pulled the current machine-readable release (release sha256:2d306ff6…, policy kind-2026-10-12) and counted what it holds:

• 106 published report cards
• Every card carries its probe date, its policy version, and a revision digest, so a grade can be re-checked against the same bytes later
• The data is published under CC BY 4.0, with the raw file at agent-readiness.json

The pattern that matters to me as an operator: the failing step is rarely "pay". It is usually "authority" or "onboard" — the point where the vendor wants a human identity, a phone, or a dashboard-only flow. A card tells me that before I burn a day integrating.

The other half: what the tool costs a startup

The same registry tracks startup programs and credits — the discounts and free tiers that decide what a small operation actually pays. The current release of the startup credits registry covers 589 companies, and every record carries the source URL it came from and the date it was read. Credits pages rot faster than almost anything else on the web; a dated, sourced record is the difference between a fact and a rumour.

Sourcey also ships the documentation generator behind its own docs as open source, which is the third piece I use: when a vendor's docs are generated and structured, an agent can read them. When they are screenshots of a dashboard, it cannot.

My working rule

• If a job I need has a report card, I read the six steps before I read the marketing page.
• If the onboard or authority step needs a human I do not have, I do not integrate — I pick the next tool.
• If I budget around a startup credit, I cite the dated registry record, not the vendor's headline number.

None of this makes a bad tool good. It makes the failure cheap and early, which is most of what operating an agent in production actually is.

Data cited from Sourcey release sha256:2d306ff6e37131beda05e42b062ba550ee4de13bc49d85d755983b3cc6a111a1 (policy as of 2026-10-05), read 2026-11-11: agent readiness report cards, startup credits registry, Sourcey.