Skip to content

Research

The research lives in the repository under docs/research/. Every claim in a brief carries a source. This page is the map.

Brief What it settles
Apple Foundation Models The runtime: 4,096-token window shared by input and output (8,192 on newer iOS 27 devices), ~70 tokens per tool definition, 3 to 5 tools max, one request in flight, overflow kills the session, guardrail and refusal errors are not retryable
React Native providers and the AI SDK The bridge: @react-native-ai/apple pre-registers tools and runs Apple’s loop inside one call, so stopWhen never fires and tool calls come back with empty ids; guided generation is real constrained decoding; the AI SDK v7 accepts V3 providers unchanged
Small-context agent patterns The harness: orchestrator sees summaries only, workers are stateless and narrow, small models fail multi-turn tool chains (8 to 56% vs 80%+ single-turn), never mix schema-constrained output with tool calls, escalate on structural signals

Eight briefs on other agent harnesses live in the helen repository. harness-survey.md synthesises them.

The lifecycle extension comparison examines Claude Code, Grok Bot, Mastra, Eve, Deep Agents, and Hermes. It informed the typed extension API.

Five briefs on what was built for small on-device models, synthesised in round2/README.md:

Rules Goliath follows maps each rule back to one of these.