AgentsAutonomy & tool use
Instinct, the text-message AI agent, reaches a $10 billion valuation as Meta's Muse and OpenAI's Dots launch rival agents
Instinct, an invite-only AI agent reached by text message, raised $1 billion at a $10 billion valuation in September as Meta's Muse and OpenAI's Dots launched similar agents. A hands-on test found it comparable on errands, though its data terms draw privacy criticism.
Reach: not measured yet
AT Pulse research deskPublished · 5 min read · AI-assisted analysisInstinct raised a $1 billion Series C at a $10 billion valuation on September 28, 2026, from Sequoia, Benchmark and Coatue. The round came one month after a Series B at $2.5 billion. The company has no app, has spent nothing on marketing and has not disclosed its underlying model. Its founder, Noah Shinn, previously worked at Sierra and is first author of the Reflexion agent research. There is no documentation that Instinct uses the Reflexion method. [2] [3] [8]
Rivals arrived quickly. Meta launched Muse on September 8. It runs on Muse Spark and is free, with paid tiers for heavier use. OpenAI launched Dots on September 29. Each Dot runs on GPT-6 Astra with its own cloud computer, and access is limited to ChatGPT Pro and Business Premium subscribers. In The Verge's hands-on, Instinct was roughly as capable as both on everyday errands. It still struggled with messy websites and missed a prerequisite class when booking swim lessons. [1] [5] [7]
Instinct's terms give it a perpetual, irrevocable licence to train on user materials, with an exception for Google Workspace data. Reports describe it sending an email without asking for confirmation and keeping Gmail data available after a user revoked access. Meta and OpenAI document safeguards: Meta uses an action monitor and requires approval for sensitive actions, and OpenAI auto-reviews account actions and keeps connections read-only during proactive research. Instinct has published no architecture or evaluation details. [4] [11] [5] [7]
How these agents will make money is still open. Analysts quoted by The Verge expect advertising or merchant deals to pay for them, and Instinct already sends users recommended products. The founder says annual transaction volume is about $1 billion, roughly half of it travel. That figure is founder-reported and described inconsistently: one outlet says the volume is approaching $1 billion, another says it has passed that mark. [1] [3] [9]
Large platforms now offer agents similar to Instinct, and Instinct's terms and reported incidents raise data-handling concerns. The likely implication, not yet proven, is that the startup must compete on distribution and trust.
Read the full assessment
Organizations should check how each agent handles revoking access, deleting data and confirming actions.
Instinct vs. Muse and Dots: dossier on the startup personal-agent race
Executive brief
Instinct is an invite-only AI agent that users reach by text message. It reached a $10 billion valuation on September 28, 2026, a month after a $2.5 billion round. It has no app, has spent nothing on marketing, and has disclosed no model details. Meta's Muse and OpenAI's Dots then launched agents that do much the same work. In The Verge's hands-on test, Instinct held its own against both on everyday chores. It also missed details, and its data terms have drawn privacy criticism. Pricing is still open. Analysts quoted by The Verge expect advertising or commerce deals to pay for these agents.
What changed and event timeline
Private beta
Instinct, founded by former Sierra employee Noah Shinn, began a private beta. It later opened an invite-only beta in August ().
$250M Series B at $2.5B
Instinct had raised $350 million in total by August. Coverage on August 24 pointed to broad data-access terms and no published technical description (;).
$1B Series C at $10B
Sequoia, Benchmark and Coatue invested, one month after the previous round. Instinct says the money goes to product work and wider access ().
OpenAI launches Dots
OpenAI announced Dots at DevDay. Each agent runs on GPT-6 Astra and gets its own cloud computer. Access is limited to ChatGPT Pro and Business Premium subscribers ().
Verge hands-on
The reviewer found Instinct roughly as capable as Muse and Dots on everyday errands. She called it effective but still limited by messy websites ().
Capabilities and access
- Instinct: Works over iMessage, WhatsApp, email or phone. It connects to Google Workspace, Slack and Notion. The underlying model is undisclosed (The Verge; MLQ).
- Muse: Runs on Muse Spark. Available in the US on iOS, Android and the web (Meta).
- Dots: Runs on GPT-6 Astra and connects to more than 4,000 apps through plugins. One Dot is included with Pro or Business Premium.
Read the full section
- Instinct: Works over iMessage, WhatsApp, email or phone. Logins go through a "secure vault." It connects to Google Workspace, Slack and Notion. It is free for invite-only users and has no app. The underlying model is undisclosed (The Verge; MLQ).
- Muse: Runs on Muse Spark. Available in the US on iOS, Android and the web (Meta).
- Dots: Runs on GPT-6 Astra and connects to more than 4,000 apps through plugins. One Dot is included with Pro or Business Premium. Users in the EEA, Switzerland and the UK are excluded at launch (SiliconANGLE).
Technical analysis for researchers and developers
Read the full section
- Shared design: All three pair a chat interface with a virtual machine plus email and calendar connectors (The Verge).
- Documented safeguards differ:
- Meta uses a separate "Sentinel" agent to watch outbound actions, requires approval before sensitive actions and keeps an audit trail (Meta).
- OpenAI auto-reviews actions that touch accounts, keeps connections read-only during proactive research, and lets a monitor pause agents (SiliconANGLE).
- Instinct has published no architecture or evaluation details (MLQ).
- Research lineage: Shinn is first author of Reflexion (2023). In that method, an agent writes reflections on feedback into episodic memory instead of updating its weights. Whether Instinct uses it is undocumented.
- No benchmark comparison of the three agents exists.
Claims and evidence
- $10B valuation and $1B round: reported by Yahoo Finance/Investing.com and Fortune.
- About $1B in annual transaction volume, roughly half of it travel: these figures come from the founder, and the wording varies.
- Instinct matches its rivals: this rests on one reviewer's tasks over a couple of weeks (The Verge).
Read the full section
- $10B valuation and $1B round: reported by Yahoo Finance/Investing.com and Fortune.
- About $1B in annual transaction volume, roughly half of it travel: these figures come from the founder, and the wording varies. Fortune says "approaching" $1 billion (Fortune). Other coverage says "surpassed" (KuCoin News). No user count has been disclosed.
- Instinct matches its rivals: this rests on one reviewer's tasks over a couple of weeks (The Verge).
- Reflexion scored 91% pass@1 on HumanEval vs. 80% for GPT-4: reported in a peer-reviewed paper (arXiv).
Context and prior work
- These are consumer versions of developer agents such as Codex and OpenClaw, which developers have used for the past year (The Verge).
- Shinn left Northeastern in 2023 to join Sierra, the enterprise-agent company co-founded by Bret Taylor (Fortune).
- Meta has since released a Muse for Small Business version.
Read the full section
- These are consumer versions of developer agents such as Codex and OpenClaw, which developers have used for the past year (The Verge).
- Shinn left Northeastern in 2023 to join Sierra, the enterprise-agent company co-founded by Bret Taylor (Fortune).
- Meta has since released a Muse for Small Business version.
Limitations, safety and contested findings
- Reliability: Instinct missed a prerequisite class while finding swim lessons. At the start it made unprompted suggestions based on the reviewer's area code (The Verge).
- Data terms: Instinct's terms grant it a perpetual, irrevocable licence to use user materials for training, with an exception for Google Workspace data (MLQ).
- Reported incidents:
Read the full section
- Reliability: Instinct missed a prerequisite class while finding swim lessons. At the start it made unprompted suggestions based on the reviewer's area code (The Verge).
- Data terms: Instinct's terms grant it a perpetual, irrevocable licence to use user materials for training, with an exception for Google Workspace data (MLQ).
- Reported incidents:
- It sent an email without asking for confirmation (MLQ).
- Emails stayed searchable after a user disconnected Google access, and Gmail records were not deleted on first request (MLQ).
- It kept summarizing Gmail hours after access was revoked, and reports describe it as open to phishing (AI Weekly).
- Instinct's own disclosure: its privacy notice says agents can take unintended actions and can meet misleading instructions hidden in third-party content.
- Muse: a reporter said Muse read private messages without being asked (Inc.).
Business and practitioner implications
- Monetization: Instinct is free for invite-only users, Muse is free with paid tiers, and Dots requires a subscription.
- Distribution: A text-only interface works anywhere messages do, including CarPlay and watches. Instinct presents this as its main difference from the big platforms.
- Governance: Before connecting work accounts, check how each agent handles revoking access, deleting data and confirming actions.
Read the full section
- Monetization: Instinct is free for invite-only users, Muse is free with paid tiers, and Dots requires a subscription. Techsponential's Avi Greengart and IDC's Greg Ireland expect agents to be paid for indirectly, through ads or deals with merchants. Instinct has already pushed "recommended products" to users (The Verge).
- Distribution: A text-only interface works anywhere messages do, including CarPlay and watches. Instinct presents this as its main difference from the big platforms.
- Governance: Before connecting work accounts, check how each agent handles revoking access, deleting data and confirming actions.
The source trail.
Sources (12)
Instinct was the buzziest AI agent around — can it survive Muse?
Article text retrieved; extracted text may omit tables or interactive elements.
theverge.com