Dust, a multiplayer AI platform for human-agent collaboration, reports roughly $10K per day saved on model spend, an 18-19% reduction in overall spend, after optimizing prompt caching with Claude. Cache reads rose from about 30% to 65% of input tokens, cutting input spend by 22%. Autonomous tool-calling depth went from 4-5 to 8+ steps per agent run with no engineering changes needed.
Dust's co-founder and CTO Stanislas Polu said: "Claude consistently stood out on the criteria that matter most: instruction-following, nuanced writing, and reliable tool use." He added: "The new Claude models didn’t just answer the question; it proactively explored adjacent information and synthesized across sources. That behavior is exactly what enterprise agents need."
Dust gives companies a shared workspace to build, deploy and manage model-agnostic AI agents connected to company knowledge, tools and workflows, without writing code. Customers can pick the model behind any Dust agent from a dropdown menu. The platform includes an LLM picker, a retrieval pipeline connecting more than 100 data sources, a framework for delegating work across sub-agents and a no-code agent builder, with permission-aware retrieval, role-based access and audit logs. Teams at companies like Datadog, Vanta and 1Password use Dust and together have deployed more than 300,000 agents.
Earlier models produced truncated or unreliable results past a handful of tool calls. Polu said: "Our existing defaults, three tools per run with a maximum of eight, were too restrictive." Once Claude could chain many steps accurately, Dust redesigned its execution loop to support up to 24 tool calls per run. Deep Research agents now orchestrate sub-agents across data warehouses, the web and internal sources, running for ten minutes or more to produce a single synthesized report. Because longer runs increased token consumption, Dust worked with Anthropic's Applied AI team on a three-tier prompt caching structure, with globally shared instructions cached for an hour and workspace and per-user context on shorter windows.
Dust also adopted Model Context Protocol, operating as both an MCP client, so its agents can reach compatible tools through one standard interface, and an MCP server that exposes its own agents and context to other MCP-aware systems. Deep research and workflow orchestration use cases have driven an average of 70% weekly active usage across customer organizations. Inside Dust, AI-written code grew from roughly 30% in early 2025 to between 60% and 90% today, depending on the engineer, with Claude Code used as a coding partner and a GitHub Action reviewing pull requests.