Anchor Action Index cut 300-task run cost from $472 to $4.87
Anchor introduces Action Index, a shared memory of proven web workflows that lets AI agents repeat successful paths and cut costs.
Talk With Tech Newsroom
The short version
Anchor unveils Action Index, a shared memory that lets AI agents reuse finished web workflows instead of starting from scratch.
In a 300-task test on Online-Mind2Web, a small model with Action Index completed tasks at an average of about $0.016 per task, while a comparable model without it would have cost much more.
A frontier model using the same setup showed 93% task completion at roughly $1.57 per task; the full 300-task run with Action Index totaled $4.87, versus $472 without it.
A paired study on 120 tasks reported sharp reductions in token use, steps, and wall-clock time, and improved performance for open-weight models.
Quick read · 2 min
Anchor rolled out Action Index, a shared memory for AI agents that lets them replay proven web workflows instead of re-solving every task from scratch. In a 300-task test, a small model using Action Index finished tasks at an average cost of about $0.016 per task, far cheaper than running without it.
The benefits extend beyond cost: a separate study showed drops in token use, fewer steps, and less time spent. Open-weight models also gained performance when using the shared memory and deterministic replay.
What this means for you: AI agents could handle more complex web work more cheaply and quickly, especially on sites without APIs. If you’re a developer, you can try Action Index today via Anchor’s MCP server or REST API. Expect memory-based workflows to become more common as AI tooling evolves.
Memory is the bottleneck for AI agents tackling web work, and Anchor is betting its answer is a shared memory of proven steps. The company released Action Index, a memory layer that lets agents replay completed workflows as deterministic code. The idea is simple: instead of re-solving every click sequence, the agent replays a known-good path and only accounts for changes on the page.
Anchor says this approach can make automated web work cheaper and faster. In a 300-task evaluation using Online-Mind2Web, a small model paired with Action Index finished tasks at an average cost of about $0.016 per task. By contrast, running the same tasks without Action Index would be far more expensive. In the same setup, a frontier model achieved 93% task completion at about $1.57 per task, and the entire 300-task run with Action Index came to $4.87 compared with $472 without it.
In a separate paired test spanning 120 tasks, Action Index reduced the 90th-percentile of token usage by 74%, cut the number of steps by 75%, and shortened wall-clock time by 61%. The benefit showed up again with open-weight models, which improved their success rates when using the shared memory and replay capability.
For everyday users, this means AI agents could handle more multi-step web tasks with less cost and less waiting. The technology could be especially helpful for tasks on sites that don’t offer APIs, where automation often has to mimic a person’s clicks and fills.
Availability is immediate: Action Index is live for all Anchor accounts, including the free tier. Developers can access it via Anchor’s MCP server endpoints and the REST API, with documentation laid out for integrating the feature into existing agent setups.
The move fits a broader trend in AI tooling toward memory and reuse. If you rely on AI agents for research, workflow automation, or customer-facing tasks, this memory layer could make automation more affordable and reliable in the months ahead.
What to do next: check whether your current agent stack can leverage MAC-compatible endpoints and review Anchor’s guidelines for shared memory usage. As with any new automation layer, keep an eye on edge cases where a previously successful workflow might require a manual nudge when a site changes.
A finished edit vanished after an AI tool gained access to production files, leaving producers to start over and pushing the show’s debut to an unspecified future date.
