Skip to content
Bareship logoBareship logoBareship

0013. Agent device tools: agent-device by default, Argent for profiling

  • Status: Accepted
  • Date: 2026-10-11

Context

Section titled “Context”

Coding agents need to run the app, tap through it, check a change on screen and profile it. Two tools do this: Callstack’s agent-device (a CLI plus skills, already vendored here as agent-device and dogfood) and Software Mansion’s Argent (an MCP server plus skills and a rule file).

A comparison on a React Native 0.87 app (Coding agents drive the simulator, Claude Code, iOS simulator, 12 scored runs) found:

  • Both solved every task: reproduce a bug, fix and verify a layout, profile and fix slow typing.
  • agent-device cost about a third as much per run ($0.41 against $1.27 median) and processed about 5× fewer tokens. It returns text snapshots with short element refs; Argent returns a screenshot and the full accessibility tree after every action, and the images stay in the conversation.
  • agent-device’s snapshot showed a [selected] state as text; Argent’s tree left it out, so its agent needed a screenshot.
  • Argent’s React and CPU profiler worked with no project changes and pointed straight at the slow code. agent-device’s React profiling installs agent-react-devtools, which edits index.js, metro.config.js and package.json; one call hung.
  • Argent’s global rule file loads into every Claude Code session, including non-mobile projects, and leaked into agent-device sessions.

The evidence is small: two runs per task, iOS only, one model, one app.

Decision

Section titled “Decision”
  • Driving the app and QA: agent-device is the default (agent-device, dogfood skills). Screenshots only when the layout itself is the question.
  • Profiling: Argent’s profiler skills (argent-react-native-profiler, argent-native-profiler), because they need no project changes. React Native DevTools (Metro j) stays the tool for humans.
  • No profiler setup in the project: don’t run agent-react-devtools init or anything else that edits index.js, metro.config.js or package.json to profile. Those files carry the release mock guard (index.demo.js, scripts/ci/check-no-mocks-in-release.sh) and the pinned-install policy.
  • Saved end-to-end tests: still Maestro (ADR 0012). Neither agent tool replaces a CI test suite.
  • Both tools are developer-machine installs, not template dependencies. Tested with agent-device 0.21.25 and Argent 0.28.0.

Consequences

Section titled “Consequences”
  • Routine agent sessions are cheaper and last longer before the context fills.
  • Two tools to keep current. Argent’s global rule file stays out of the global Claude rules; argent update copies it back, so move it again after each update.
  • Both tools read the accessibility tree. Interactive elements need aria-label, and toggles and tabs need aria-checked / aria-selected, or agents fall back to screenshots and coordinates.
  • The comparison covered iOS only. Confirm agent-device on the Android emulator before relying on it there.

Alternatives considered

Section titled “Alternatives considered”
  • Argent for everything: about 3× the cost per routine run, and its global rule loads into every session.
  • agent-device for everything: its React profiling edits project files this template guards, and was the slowest path in the comparison.
  • Both loaded by default: conflicting instructions and wasted context, as the comparison found.