Asana

Overview

Asana describes bringing agent workflow automation to its platform through StackAI. Its browser-agent study evaluates cost and runtime while checking answers against an independently prepared reference. (source)

Key People

• Frank Hidalgo authored the browser-agent study; Arnab Bose is identified as Asana's CPO. (source)

Models & Products

• StackAI receives the browser-navigation changes; Command records requests, traces and results and connects findings to reviewed changes. The study uses Astra in Codex to conduct experiments and GPT-6.1 Sol as one evaluated model. (source)

Recent Activity

• 2026-10-08: tested six history/caching policies at two character budgets across four models, with three runs per condition: 144 runs, plus a 12-run follow-up. On the original Model B, optimization reduced cost 29x and runtime 4x; switching to the optimized GPT-6.1 Sol workflow produced the headline 76x cost and 5x runtime reductions against that original baseline. Capped baseline runs make those reductions lower bounds. (source)
• The fix caches growing history, retains screenshots until 20 accumulate before reducing to 1, and increases the text budget from 120,000 to 480,000 characters. With the larger budget, every Model C and Sol run answered; the smaller budget yielded none of 18 Model C runs and 3 Sol runs. (source)

Strategic Position

Interpretation: this is an example of making agent deployment economical through history management and measurement. The study explicitly supports broad patterns rather than small differences between conditions, and says compaction was outside its scope. It does not establish a universal model ranking. (source)

Related

• Agents (LLM Agents) · Eval Harness Configuration · OpenAI — connections through agent orchestration and the evaluated models. (source)