Atlas: agents explore once, every agent after takes the highway
A shared map of websites in HydraDB. Browser agents plan routes with one shortest-path query and call the model only where the map ends.
Illustrative trace of the idea, not benchmark data. See it for real in the live demo.
How it works
A task is a route. Atlas finds the route on the map and only leaves it where the map ends.
Locate
Fingerprint the current page and find which node of the map the agent is standing on.
Plan
One shortest-path query in HydraDB returns the cheapest, most reliable route to the goal.
Drive
Replay each learned action in the browser. No model in the loop.
Explore and learn
Where the map ends, the model takes one step. The result is written back, so the next agent drives it.
Why a graph
Routes are paths. In a graph database, planning is a query instead of a prompt.
Composition
New tasks reuse learned edges. A task nobody ran before is stitched from roads other tasks already paved.
Self-repair
Edge reliability is tracked. When a site changes, failing edges are flagged and the model re-explores only that stretch.
Shared map
One map per site, shared by every agent. The first agent pays for exploration; the rest ride free.
Memory for agents: skills, tools, entities
Atlas keeps skills, tool definitions and resolved entities in the graph, and pulls only what the current task needs instead of filling the context window.
Skills
Frequent sub-paths are distilled into named skills that agents can call whole.
Tool definitions
The actions available on each page live on its node and are retrieved when the agent arrives there.
Resolved entities
Orders, products and accounts already resolved are recalled rather than searched for again.
Use it from your agent (MCP)
Atlas speaks the Model Context Protocol. Point any MCP client at the server and the map is yours.
{
"mcpServers": {
"atlas": { "command": "python", "args": ["-m", "atlas.mcp"] }
}
}- atlas_find_skills
- Search learned skills and routes relevant to a goal.
- atlas_run
- Run a whole task: plan on the map, drive it, explore only where it ends.
- atlas_act
- Take a single mapped action from the current page.
- atlas_recall
- Recall resolved entities and facts from earlier runs.
Results
Model calls, steps and wall time: first run versus later runs.