Composite

Composite's whole bet is that the hardest part of browser automation is not the clicking — it is knowing what you wanted done. Everything on the site points back to that. The intent-detection research published September 21, 2026 does not claim perfect grounding; it says grounding was the easy part and inferring intent was the hard part, and then reports 73% correct when it acts at its most selective checkpoint, delivered in under 400ms per prediction to beat human reaction time. That number is the one to hold onto, because it is the same shape as the Composite-Bench leaderboard the company published on July 24, 2026: their 27B open-weights base, fine-tuned and then trained with RL, sitting between GLM-5.2 at 63% and Claude Opus 4.8 at 73%, with GPT-5.6 Sol at 18% and Grok-4.5 at 9% on the same 100 hidden tasks. A vendor publishing its own 73% rather than a round 90% is telling you something useful.\n\nStrengths. Setup is genuinely minimal: install for macOS, press Cmd+Shift+Space on any website, and the agent works on that site immediately — no integrations to connect, no workflows to build, nothing for IT to approve. It runs in your existing browser where your logins, tabs and bookmarks already live, so it inherits your authenticated sessions rather than asking you to re-authenticate into a hosted environment. Background tab execution means the work happens while you keep using the foreground window; watch mode lets you step through it instead. Execution stays local on your Mac with zero data retention with AI sub-vendors, which is a materially different posture from hosted RPA that ships your session to a cloud. It learns from how you work and then proactively suggests tasks, so the value curve bends upward with use rather than staying flat.\n\nWeaknesses. It is macOS-only — Windows and Linux users cannot run it at all. The published capability is 73% at its most selective checkpoint, which is strong for long-horizon browser work but means you should still check outputs on anything multi-step. Anti-bot pages such as Cloudflare challenges can break a run mid-flight. The public site describes the product for individuals and says nothing about shared workflows, audit logs or approval chains, so teams with governance requirements should not treat it as infrastructure. And high-volume structured extraction into CSV or JSON is better served by a purpose-built scraping tool.\n\nWhere it fits. Solo professionals — accountants, recruiters, analysts, salespeople, freelancers and consultants — with recurring weekly browser busywork across many sites: submitting expenses through finance portals, pulling vendor information into comparison sheets, assembling sales updates from several dashboards, tracking postings across job boards, collecting competitor pricing on a daily loop. It also fits people who want to evaluate computer-use agents on their own real tasks before committing budget, since the install cost is one shortcut.\n\nWhere it doesn't. Teams