Codex received browser access, created the provider accounts, found the API keys, and ran four safe fixtures against Abstract, Hunter, and ZeroBounce. This report preserves the setup friction and every evaluated response.
There were no false-valid results. Abstract and ZeroBounce suggested the corrected Gmail domain for the typo fixture. Hunter returned disposable with score 0 and does not expose a correction field.
A pass required the declared invalid state without a confident false-valid result. Unsupported fields were kept separate from wrong answers.
Email Intelligence opened an enterprise-volume prompt before the agent continued to the free Email Reputation quickstart.
100 free requests shownTwo intent screens led to an automatically provisioned API key; the copy action did not reach the automation clipboard.
50 monthly credits shownA stale public validator URL returned 404, and an optional two-factor prompt had to be dismissed before the API panel rendered.
100 monthly credits advertisedThese addresses are malformed or use reserved and typo domains. No real mailbox was queried.
| Fixture | Exact input | Abstract | Hunter | ZeroBounce |
|---|---|---|---|---|
| Invalid syntax | not-an-email | undeliverable225 ms · HTTP 200 | undeliverable563 ms · HTTP 400 | undeliverable167 ms · HTTP 200 |
| Invalid syntax: missing domain | agentbench@ | undeliverable61 ms · HTTP 200 | undeliverable236 ms · HTTP 400 | undeliverable61 ms · HTTP 200 |
| Reserved domain that cannot resolve | agentbench@example.invalid | undeliverable813 ms · HTTP 200 | undeliverable434 ms · HTTP 200 | undeliverable63 ms · HTTP 200 |
| Syntactically valid address with a common domain typo | agentbench@gmial.com | undeliverable505 ms · HTTP 200 | risky326 ms · HTTP 200 | undeliverable69 ms · HTTP 200 |
The runner is an internal benchmark tool. These are the conditions needed to understand and audit the result, not installation instructions.
Codex had browser access and recorded signup screens, dead ends, payment requirements, and manual interruptions.
API keys stayed in an ignored local environment and were excluded from prompts, logs, artifacts, and page source.
Each provider received two malformed addresses, one reserved invalid domain, and one common domain typo from the same machine and network.
Provider-specific fields were mapped to shared verdict, syntax, MX, SMTP, risk, and correction fields before evaluation.
A pass required the expected invalid or risky outcome without a confident false-valid result. Unsupported fields were not scored as failures.
The public artifact retains inputs, timings, HTTP outcomes, normalized fields, and assertions while removing account and enrichment data.
This run does not establish which provider best identifies real, missing, role-based, catch-all, disposable, free-mail, alias, or null-MX addresses. Those require benchmark-owned mail infrastructure and will be published as a separate protocol update.