Company profile · 2 min read
Oasis: Shared context for people and agents
Oasis is a shared room where humans and agents work on the same task state.
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
Published · Updated
Oasis is a shared room where humans and agents work on the same task state.
What it does
The homepage describes rooms as group chats between people and agents, with everyone able to work in the same document, spreadsheet, or codebase. Oasis also stores living memory of decisions, tasks, and lessons so a new agent can continue without a cold start. It advertises automatic routing, shared agents, and fleets of specialized agents.
The agents page adds an important deployment choice. Agents created in Oasis run in its cloud by default, while local agents can be connected through the Mac app. Agents can message one another, hand work off, and work with services and apps from a shared room. That makes the product’s real question less “which model is smartest?” and more “which agent is allowed to change which shared state?”
Buyer and task
The buyer is a team coordinating several agents across a workflow where context and handoff are expensive. Start with one room, one document or codebase, and two roles—for example, a researcher handing evidence to a writer. Define which actions require a human and which agent owns the final state.
Oasis says teams own their knowledge while its living memory helps route work automatically. The public pages do not establish reliability, agent autonomy, or organizational outcomes. A team should also decide whether a task may use a cloud agent, a local agent, or both.
The first test
Run a bounded task with a cloud agent and a connected local agent. Record the room state, each handoff, every file change, and the final human approval. Swap one agent midway and check whether the successor receives the intended memory without inheriting an obsolete instruction. Then test a deliberately restricted connection.
The privacy policy says Google data is used only for the user-facing features authorized and is not used to train generalized models. Verify that behavior, retention, and deletion terms in the customer agreement before connecting sensitive work.
Sources checked
official Speedrun profile · company homepage · agents · privacy policy
Sources checked — September 20, 2026.
About the author
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
I cofound Lazyweb and publish Mudpie. This is an owner-written publication, not an independent testing organization. Research notes distinguish observations, sourced reporting and editorial judgment.
First1000 ↗ · X ↗