Anchor, the browser infrastructure company for AI agents, has launched a tool it says can sharply cut what it costs an agent to finish work on the web. The product, called Action Index, is a shared memory of web workflows that agents have already completed successfully. Rather than working out every click from scratch, an agent on Anchor replays a proven path as deterministic code. The model steps in only where a page has changed.
The company tested the approach on 300 tasks from the Online-Mind2Web benchmark. A small model equipped with Action Index completed 93.3% of those tasks at an average cost of $0.016 per task. A frontier model completed 93% of them at $1.57 per task. Anchor said the full 300-task run cost $4.87 with Action Index, compared with $472. The same small model, without Action Index, completed 62.7% of the tasks.
A separate paired study covered 120 tasks. Anchor reported that Action Index reduced 90th-percentile token use by 74%, steps by 75% and wall-clock time by 61%. The gains were not limited to Anchor's own models. Two leading open-weight models rose from 67.7% to 88.7% and from 58.7% to 84.7% when they used the system.
Idan Raman, co-founder and chief executive of Anchor, framed the launch around memory rather than raw model capability. “Agents today pay to re-solve problems that are already solved,” Raman said. “The model isn’t the bottleneck for most web work anymore. Memory is. Action Index means the hundredth run of a workflow costs a fraction of the first, and every successful run makes the next one cheaper.”
That idea of cost falling with repetition is the core of the pitch. Each completed workflow becomes a reusable asset that later agents can draw on, so the system gets cheaper the more it is used. Anchor described the tool as a way to address the expense of running agents that interact with live websites, where a model that reasons through each step can burn through large numbers of tokens.
Action Index is available now as tools on Anchor's MCP server and through its REST API. The company said it works with any MCP-compatible agent or model provider, which means developers do not have to switch to Anchor's models to use it. The MCP endpoint is listed as https://api.anchorbrowser.io/v1/mcp, and documentation is available at https://docs.anchorbrowser.io/advanced/mcp.
Anchor describes its core business as giving agents secure, reliable cloud browsers that can log in to, navigate and complete tasks on any website, including sites that offer no API. That positioning matters for Action Index, since many of the workflows it stores involve pages built for people rather than software. The launch is aimed at developers and companies building agents that have to operate those pages.
The results come from Anchor's own benchmarks and have not been independently verified. The company's figures are most meaningful when read against the specific tasks and models it chose to test, and the comparisons are between Anchor's configurations and baseline runs it ran itself. Anchor has not said how many workflows the shared memory currently holds or how many accounts are using the feature so far.
Dieser Artikel wurde mit Unterstützung von KI verfasst.
News Factory APP - agentische News für besseres SEO & AEO.