Anchor Action Index Cuts Computer-Use Agent Token Costs by 100x
Computer-use agents are expensive because they reason through every task from scratch. Anchor, the browser infrastructure company, launched Action Index on October 15 to fix that.
Action Index is a shared repository of web workflows that agents have already completed successfully. When a new task matches a stored workflow, the agent skips the model loop and replays the proven sequence as deterministic code. No re-reasoning. No re-clicking. The model only gets involved when the task is genuinely novel.
The cost difference is 100x. Reasoning through a web task requires multiple model calls — planning, observation, correction — each generating tokens. Replay runs the recorded code path directly, bypassing the model entirely.
What This Means for Computer-Use Economics
Computer-use has a well-documented cost problem. Agents that reason visually through browser tasks generate far more tokens than equivalent structured-API calls, and they fail more often. A 2026 study pegged the cost difference between computer-use and structured APIs at 45x for identical tasks.
Action Index attacks the problem from a different direction: rather than making the model faster or cheaper, it removes the model from the critical path for known workflows. The first agent to complete a task pays the full reasoning cost. Every subsequent agent that hits the same workflow pays the replay cost, which is effectively zero in model-call terms.
This is a shared commons model. Anchor accumulates workflow solutions across all tenants, which means the index improves with scale. A workflow solved once for any customer becomes available to all.
Architecture and Scope
Action Index covers web workflows on Anchor’s browser infrastructure. The system identifies when an incoming agent task matches a stored workflow with sufficient confidence, substitutes the deterministic replay path, and falls back to live model reasoning when no match is found or when the live environment has diverged from the recorded state.
Anchor detects divergence when DOM selectors or page states no longer match the recorded baseline, triggering a step-level model fallback for that specific step before resuming deterministic replay. Actions that persistently fail are retired or re-verified. The system is designed to degrade gracefully rather than fail hard when the live environment has changed.
Why It Matters
Computer-use is where the next generation of AI agents lives. Tasks requiring navigation of arbitrary web UIs — booking, research, form filing, data extraction — cannot be handled by structured APIs. Every major AI lab has a computer-use capability; the question is whether the economics are viable at scale.
Action Index is a credible attempt to change the denominator. If the index grows large enough to cover the majority of common workflows, the effective cost of computer-use drops to the replay floor for most tasks. That changes the build-vs-buy calculation for any team using agents against production web applications.