Perplexity gives GPT-6 Astra end-to-end production access
OpenAI's customer story skips the numbers that matter. The lesson for your own agents isn't the trust — it's matching autonomy to what you can cheaply undo.
Copy markdown
What Perplexity says it handed the model
Cofounder Johnny Ho says Perplexity now lets Astra craft communications, edit real-world systems, and monitor production software — and "check in on it much less frequently than previous generations." One cited use: Astra stands in for external services, generating realistic API responses so apps can be tested end-to-end.
The numbers OpenAI didn't publish
The post ships zero metrics — no baseline, no error rate, no scale. "Much less frequent" than what: hourly to daily, every commit to every deploy? And "monitor production" is left ambiguous between passive alerting and active remediation with write access — the gap between a log-reader and a finger on the deploy button.
Three tasks, three risk curves
The jobs aren't equal. Comms are naturally gated by human review; software edits are recoverable if you have code review and rollback; autonomous production monitoring is the danger zone precisely because nobody says what's automated. Bucket an agent's actions by blast radius, not by a blanket "trust level."
The takeaway for your agents: autonomy = reversibility
Copy the principle, not the headline: let agents run free on anything a human can cheaply undo, and keep a hard gate on anything expensive or impossible to reverse. Fewer check-ins isn't a feature you buy from a smarter model — you earn it from your own guardrails: code review, rollback, scoped write access.