Apple on Monday introduced AgentKit, a software framework that runs multi-step artificial intelligence tasks entirely on iPhone and Mac hardware without routing prompts through cloud servers. The announcement, made at the company's annual developer conference in Cupertino, Calif., marks Apple's most direct response yet to enterprise buyers who have blocked cloud-based assistants over data residency rules.
AgentKit allows third-party apps to chain on-device model calls — search, calendar updates, document summarization — within a single user-approved workflow. Apple executives said the system processes personal data only in encrypted local memory and deletes intermediate states when a task completes. For regulated industries, that architecture removes a primary objection to generative AI adoption.
What Apple Shipped
The framework builds on Apple Intelligence models released in late 2025. Developers gain APIs to register tools — email clients, file browsers, health records — that agents may invoke after explicit user consent for each session. Apple demonstrated a travel rebooking scenario: the agent read a cancellation email, searched alternate flights, and presented options without transmitting mailbox contents to external servers.
Technical briefings reviewed by Credence Wire show inference latency averaging 1.2 seconds per step on iPhone 17 Pro hardware, using a 7-billion-parameter model optimized for Apple's neural engine. The company did not disclose parameter counts for larger on-device models under development.
Enterprise administrators receive mobile device management controls to disable agent tool access by app category. Banks and hospital systems piloting the beta reported they could permit summarization on internal documents while blocking web browsing tools entirely.
Competitive Context
Microsoft and Google continue to route most Copilot and Gemini features through cloud infrastructure, though both have expanded offline subsets. Apple's bet is that privacy guarantees will unlock deployments in European financial services and U.S. health care, where general counsel offices have slowed cloud assistant rollouts.
Samsung and Huawei ship on-device assistants in Asia, but neither offers a third-party agent framework with comparable tool-registration APIs. Analysts at Canalys said Apple's timing aligns with EU Data Act enforcement windows that penalize unclear cross-border data flows.
Developer Reception
Attendees at WWDC described AgentKit documentation as clearer than last year's Apple Intelligence rollout, which frustrated teams with shifting beta APIs. Still, several engineers noted that complex agents requiring reasoning over large corpora will remain cloud-bound — local models cannot yet match frontier systems on open-ended research tasks.
Adobe, SAP, and Salesforce announced same-day integrations for document and CRM workflows. Each stressed that customer data stays on-device for defined actions, though optional cloud escalation remains for tasks users approve individually.
Regulatory and Labor Questions
European regulators welcomed on-device processing as compatible with GDPR minimization principles. U.S. lawmakers focused on whether agent actions constitute automated decision-making under pending AI governance bills. Apple published a white paper describing human confirmation requirements for payments, account deletion, and outbound communications.
App-store workers' advocates asked whether agent-driven automation would reduce demand for human virtual assistants and customer-support contractors. Apple declined to project employment effects, saying the framework targets task completion rather than job replacement.
What Comes Next
AgentKit enters public beta in July and general availability with iOS 20 and macOS 16 in September. Enterprise customers on Apple Business Manager can enroll in a managed preview next month. Industry watchers will track whether on-device constraints limit adoption or whether privacy assurances accelerate Fortune 500 deployments that cloud assistants have failed to secure.
