react-brainAI-assisted development — agent skills, MCP tooling & device automation
draftedconfidence: lowearly: 1/12 of this tier graded so far (0 overturned) — the public scorecard →
recommendation
Give a coding agent two things: RN KNOWLEDGE — a curated skills pack (Callstack agent-skills is the broadest; add Margelo react-native-skills for camera/native depth) — and EYES+HANDS on the running app — Maestro MCP if you already test with Maestro (see RB-E-TESTING), else agent-device (cross-platform) or Argent (iOS debug/profile depth). Add Reactotron MCP when the agent needs runtime state. Fast-moving (confidence: low) — most of these tools are months old; trial on one workflow before standardizing.
- agent writes RN code → install a skills pack (agent-skills / react-native-skills) so it stops generating stale patterns
- agent must run and SEE the app (closed verify loop) → agent-device (cross-platform) or Argent (iOS debugger/profiler depth)
- agent-generated E2E tests → Maestro MCP (owned by RB-E-TESTING)
- agent needs runtime state / network inspection → Reactotron MCP
- how to WORK with agents (TDD, review discipline, cognitive debt) → agentic-engineering-patterns
Low confidence — fast-moving or lightly-vetted domain: treat the pick as a vetted lead and prototype before committing.
Options & tradeoffs
| option | tradeoff | evidence |
|---|---|---|
| agent-skills (Callstack) | curated skill packs for coding agents — RN best practices, upgrade playbook, brownfield migration, library scaffolding; installs into Claude Code / Cursor / Codex / Gemini CLI; the broadest pack (~1.5k★) | |
| react-native-skills (Margelo) | RN-focused skills (VisionCamera and native-module depth) via `npx skills add margelo/react-native-skills`; younger/smaller than Callstack's | |
| agent-device (Callstack) | agent-native CLI: AI agents drive real iOS/Android/tvOS/macOS/web devices & simulators — a11y snapshots, actions, replayable scripts; v0.19.x, iterating fast | 91k/wk · ships in 1/34 |
| Argent (Software Mansion) | MCP server + skills: simulator control, debugger attach (React tree + JS), console/HTTP monitoring, correlated React+iOS profiling; 0.15 (2026-07, verified vs npm) extends beyond iOS — tvOS/Android TV/Vega OS, cloud agents for autonomous bug repro, Argent Lens design review; also listed in RB-E-TESTING | 45k/wk |
| Maestro MCP / Maestro Viewer | agents drive devices and GENERATE E2E flows on the tool the ecosystem already tests with — see RB-E-TESTING (which owns the testing depth) | |
| Reactotron MCP | the classic RN debugger exposes runtime state/network to coding agents via MCP (reactotron-core-server 3.3.0, 2026-05) | 307k/wk · ships in 4/34 |
| Expo AI tooling | SDK 55+ ships agent-focused tooling (Expo skills, MCP integration) for Claude/Cursor/Codex as part of the framework; Expo Agent (waitlist beta, 2026-03) is their first-party app-building agent — track, don't bet | |
| TanStack Intent (ALPHA) | a DISTRIBUTION mechanism, not a pack: agent skills ship INSIDE npm packages and are discovered via the node_modules dependency graph, version-pinned with staleness detection; rolling out in TanStack DB first | |
| React Native Evals (Callstack) | open-source benchmark suite for WHICH MODEL writes good RN code — task dataset by category/library, TS+Bun runner, judge + methodology whitepaper; the model-selection layer under all the tools above |
npm weekly downloads (from the corpus's last signals run): agent-device 91k · reactotron-react-native 307k · @swmansion/argent 45k
Verified notes
NEW entry (2026-07-09): the dominant recurring theme of Native Weekly issues 11–16 (Jan–Jun 2026) — every single issue shipped agent-tooling news (agent-device ×3, two skills packs, Reactotron MCP, Argent, Expo SDK 55/56 AI tooling, Xcode 26.3 'agentic coding') — and the corpus had no home for the SELECTION question 'which agent tooling for RN development'. Distinct from RB-E-AI-UI (AI in the product) and RB-E-ONDEVICE-AI (models on device): this is AI at DEV time. Device-driving TESTING tools (Maestro MCP, Argent, Radon IDE) stay owned by RB-E-TESTING; this entry is the cross-tool map. All options fetch-verified 2026-07-09 (GitHub repos, npm, SWM blog). Everything here is pre-1.0-culture — pin versions, expect churn.
Canonical reading
Editorial annotations on why each piece matters — the articles themselves are the originals; read them there.
The first named production-adopter case study for agent-device: agents run Sentry-span measurement loops across branches and drive the React profiler mid-session via react-devtools integration, returning named components and render counts as bug evidence. What agent-in-the-debug-loop looks like at a real company, beyond vendor demos.
The clearest articulation of the closed loop this entry is about: the agent that writes the code also boots the simulator, drives the UI, attaches a debugger, and profiles React+iOS in the same session — with production numbers (~50% re-render reduction in a banking app) rather than demos.
Sources
- github.com/callstackincubator/agent-skills
- github.com/margelo/react-native-skills
- github.com/infinitered/reactotron/releases/ta…
- github.com/callstack/agent-device
- tanstack.com/blog/from-docs-to-agents
- callstack.com/blog/announcing-react-native-evals
Depth (in-domain rules) is owned by the agentic-engineering-patterns skill — this entry is selection breadth.
Related in ai: ondevice-ai · ai-ui