Asana makes browser agent 76x cheaper with GPT‑6 Astra

Asana's browser agent work is notable because the gains did not come from switching to a cheaper model. The original agent cached fixed instructions and tool definitions but resent its growing history of page text and screenshots at full price on every request, and it trimmed or dropped that history at nearly every step. Those edits changed the history continuously, which meant caching the history alone would not have worked, while losing facts could force the agent to revisit pages it had already read.

Hidalgo selected three fixes to test: extending caching to browsing history, increasing retained text, and removing screenshots in batches. GPT‑6 Astra first refactored the code so many workflows could run in parallel with separate settings, then ran a 144-run study across GPT‑6.1 Sol and three other frontier models, recording requests, traces and results in Asana's Command platform so the team could review the study afterward and turn findings into tickets and pull requests.

The broader pattern Asana describes is a human setting direction while an agent runs experiments — Hidalgo says he would set a goal before going to bed and review results in the morning, and Asana is now using GPT‑6 Astra in Codex to test product features before release. Whether that approach translates to other teams depends on how much of their workflow can be parallelized and recorded this way.