Dust in autumn 2026: a $40M Series B, Sidekick, agents built from conversations, and the browser

In April we wrote that Dust would live or die by access control. Six months on, the company has raised a $40 million Series B, reframed itself around "multiplayer AI" (people and agents in shared workspaces), and shipped a run of features that make building an agent a conversation rather than a form. This is the autumn update for UK buyers who shortlisted Dust and want to know what changed.

What Dust is (still)

Dust is an enterprise platform for building AI agents grounded in company knowledge (Notion, Slack, Google Drive, GitHub, HubSpot and more), with a permissioning layer that separates who can access data from who can use an agent. It is model-agnostic across OpenAI, Anthropic, Google and Mistral, and sells on SSO, SCIM, audit and SOC 2 rather than on any one model.

What changed since April

The money and the pitch

On 18 May Dust announced a $40M Series B led by Abstract and Sequoia, with Snowflake Ventures and Datadog participating, taking total funding past $60M. The framing is "multiplayer AI": agents as participants in team workspaces, not private copilots. For a buyer, the practical signal is runway and a product direction that favours shared, governed agents over individual chat.

Building agents got conversational

• Sidekick (April) is a co-pilot inside the Agent Builder that drafts instructions and proposes improvements as reviewable changes
• Build agents and skills from conversations (5 October): while chatting with an agent, you can turn what worked into a new agent or skill, with review cards to accept or decline each change before it lands. This is the feature that moves agent-building from "the one person who knows the builder" to anyone who had a useful conversation

Models and reach

• Claude Sonnet 5 and GPT-5.6 are available in the Agent Builder, alongside earlier Claude Sonnet and Opus 4.6 options
• Voice expanded to 21 languages
• Agents can write to Google Drive, not just read it
• A Firefox extension joined the Chrome one, and both now give agents awareness of the page you are on
• A "Go Deep" tool hands complex tasks to deeper analysis, and AGENTS.md management in Pod settings lets technical teams version agent instructions as files
• The product moved to a single app.dust.tt address

Partnerships

Dust named AI Partners an official partner on 23 September and announced a partnership with Mantu on 15 October. For UK buyers, that is an early sign of an implementation ecosystem forming around the product, which matters when you want someone other than the vendor to run the rollout.

Who should care, who should wait

Strong fit

• Mid-sized organisations with knowledge spread across Notion, Drive and Slack who want many small, governed agents rather than one big assistant
• Teams that have already hit the "who built this agent and what can it see" problem with ungoverned tools
• Organisations with a model-choice policy (for example, Anthropic for some workloads, OpenAI for others) who want one builder across them

Wait or constrain

• Companies whose data lives mostly in systems Dust does not connect to yet; check the connector list before the demo
• Teams that need deterministic, scheduled automation with retries first. That is an n8n job; Dust is the knowledge-and-reasoning layer that sits beside it
• Anyone who cannot resource the permissioning work. Dust's model is only as good as the data-access rules you set up

Evaluation checklist for UK operators

• Map data access before agents. Which spaces can which groups see? Dust enforces it; you define it
• Pick three agents, not thirty. A sales-call prep agent, a policy Q&A agent and a weekly ops digest cover most of the early value
• Test the browser extension on your real tools. Page awareness is powerful and raises the same questions as any browser-using agent; see Agents that click
• Decide who may publish. Agents built from conversations are easy to create; agree who reviews before an agent reaches the whole company
• Model policy. Set which models are allowed per space and why, and write it down for procurement
• Exit. AGENTS.md and exportable instructions help; check what else you can take with you

Dust vs the alternatives

Bottom line

Dust spent 2026 making the thing we worried about in April (permissioning) the centre of its pitch and then lowering the cost of building agents on top of it. Series B money, Sidekick, agents-from-conversations and browser awareness add up to a product that is easier to adopt widely and therefore easier to adopt badly. Set the access model first, then let people build.

Related reading

• Dust and the permissioning problem
• Your agents need logins, not API keys
• Enterprise AI agents that survive production
• Dust on the marketplace
• Browse AI agents

Sources: Dust Series B announcement · Dust product updates · Dust changelog · Build agents and skills from conversations (docs changelog PR) · Yahoo Finance on the raise

Erhan Timur, Founder, Digital by Default