SwiftAgent 1.0.0-rc.3 is out — and today we’re also open-sourcing OtohaChat, a real app built on top of it. We’ve been exploring a simple question: What does an AI agent runtime look like if Swift is the foundation — not just the UI language? 🧵 SwiftAgent is a provider-neutral AI harness written in Swift 6.4. Its job is not to be another giant “AI framework.” It provides the execution layer: Agent → Session → Run → Model → Tool → Result → Journal with cancellation, durable recovery, model switching, tool receipts, usage accounting, and explicit host-owned permissions. RC3 supports OpenAI Responses, Anthropic, DeepSeek, local / compatible Responses APIs, Apple Foundation Models, and an experimental Private Cloud Compute path. The important part is that providers do not own your tools. Your app does. The model can propose an action, but authorization, execution, evidence, receipts, persistence, and recovery remain under the Host’s control. That’s the design principle behind SwiftAgent. To show what this looks like outside of tests, we open-sourced OtohaChat: github.com/OtohaCo/OtohaChat It’s a native SwiftUI chat app using SwiftAgent RC3, with: • Apple Foundation Models • Apple Private Cloud Compute • OpenAI / Anthropic / DeepSeek • local models and compatible gateways • model switching • reasoning controls • durable conversations • host tools • .agents/skills • Keychain-backed provider configuration There’s also a public TestFlight for iOS and macOS: testflight.apple.com/join/BC… OtohaChat isn’t just a UI demo. It exposes the actual harness: sessions, streaming runs, tools, journals, model transitions, host permissions, and execution results. The same runtime is being used inside Otoha AI Player. Why Swift? Because Swift is becoming interesting far beyond “write an iPhone app.” With Swift 6.4, the same language is increasingly relevant across: Apple platforms Linux servers Android and even embedded systems. AI agents need more than model APIs. They need runtimes, tools, state, permissions, persistence, and interfaces to the real world. We think Swift can be a very good language for that layer: native performance, strong types, structured concurrency, memory safety, and excellent C interoperability. Our longer-term goal with SwiftAgent is therefore bigger than chat: make Swift a practical language for building AI-powered software from the user interface all the way down to services — and eventually devices. SwiftAgent RC3: github.com/OtohaCo/SwiftAgen… OtohaChat: github.com/OtohaCo/OtohaChat Both are open source. If you’re building AI software in Swift, we’d love for you to try them, break them, and tell us what the ecosystem still needs. Otoha technology otoha.co/technology/ @SwiftLang @AppleMusic @OpenAIDevs @ClaudeDevs @Apple @typesafeai @deepseek_ai @OpenAI @AnthropicAI

Updated SwiftAgent, and its cache hit rate improved.

Otoha AI harness scored 80% in our internal Frontier Harness tests on GPT-6.1-Sol, above Codex and Claude on Kimi K3. It also reached 60% on GPT-6-Luna. Cache collision handling remains weak and is still being improved.

SwiftAgent is designed for trusted enterprise applications, so auditable and traceable records are essential. What other capabilities do you think are needed?

SwiftAgent 1.0.0-rc.5 is out 🎉 (SDK prerelease) • Opt-in bounded replanning • Tools defined at runtime • Route continuation isolation • Journal durability fixes Otoha AI "おとはちゃん", built on this harness, is coming soon. Stay tuned! github.com/OtohaCo/SwiftAgen…

Do you want to listen to podcasts from other countries? In Otoha Player, you can import Apple podcasts and RSS dot com from whole world, as well as AI music you make yourself on Suno.

SwiftAgent 1.0.0-rc.2 is out. SwiftAgent is a provider-neutral Agent runtime for Swift, built around one idea: keep the Agent loop small, predictable, and trustworthy. RC2 adds stronger session/runtime guarantees, durable tool execution, multi-provider support, model discovery, run-level model switching, usage tracking, and safer context handling. The goal isn’t to build another “AI framework that does everything.” It’s to give Swift apps a solid execution layer for agents — whether you’re building a coding tool, a desktop assistant, a music DJ, an automation app, or an on-device / cloud hybrid experience. OpenAI, Anthropic, DeepSeek, local models, Apple Foundation Models — one runtime, host-owned tools and state. SwiftAgent RC2: github.com/OtohaCo/SwiftAgen…

Otohaは #Suno の #AI楽曲 を直接インポートできて、本当に素晴らしい音質です💋AIのパワーがすごい

the cruel thing about learning a language is knowing every word on the page and none of them in the air.