Dust powers enterprise AI agents with Claude Platform API

Title

Dust powers enterprise AI agents with Claude Platform API

Content

Dust, a multiplayer AI platform for human-agent collaboration, deployed Claude via the Anthropic API across its agent infrastructure, saving roughly $10,000 per day on model spend — an 18-19% overall reduction — after optimizing prompt caching. “You can’t save time with AI you don’t trust. Our goal was to make AI agents accurate and capable enough that enterprise teams trust them to do real work, not just answer questions,” said Stanislas Polu, Co-founder and CTO of Dust. On the model choice he added: “Claude consistently stood out on the criteria that matter most: instruction-following, nuanced writing, and reliable tool use,” and later: “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’s premise is that the people closest to the work — the RevOps lead automating deal prep, the chief of staff rebuilding onboarding, the support manager turning ticket routing into a system — should be able to build the agents they need without writing code. Earlier Claude generations produced truncated or unreliable results past a handful of tool calls, and Dust’s existing defaults (three tools per run, max eight) were too restrictive: agents would hit the tool limit and stop mid-task. As Claude’s agentic capabilities improved, Dust redesigned its execution loop to support up to 24 tool calls per run. Autonomous tool-calling depth increased from 4-5 to 8+ steps per agent run with no engineering changes, and inside Dust itself AI-written code grew from roughly 30% in early 2025 to between 60% and 90% today. The platform runs Claude via the Anthropic API and adopted Model Context Protocol (MCP) as a standard integration layer, operating as both client and server. A retrieval pipeline connects more than 100 data sources, a sub-agent framework delegates complex work, and a no-code agent builder sits on top. Dust worked with Anthropic’s Applied AI team on a three-tier prompt caching structure: globally shared instructions cached for one hour, with workspace and per-user context on shorter windows. Cache reads doubled from about 30% to 65% of input tokens, cutting input spend by 22%. Deep Research agents now orchestrate sub-agents across data warehouses, the web, and internal sources, running for 10+ minutes to produce a single synthesized report. Customers including Datadog, Vanta, and 1Password have deployed more than 300,000 agents on Dust, with 70% weekly active usage across customer organizations. “The role of an engineer at Dust is evolving from writing code to directing, reviewing, and orchestrating AI-generated output,” Polu said — a shift the company is using Claude Code to deepen, both as a day-to-day coding partner and as a GitHub Action reviewing pull requests.