Chrome Controlla: open-source control layer makes browser agents verify whether actions actually happened
Swimming_Ask3859 · reddit · 2026-11-02
A Reddit developer argues agents are getting smarter faster than they're getting reliable: the real bottleneck is the execution layer, where uncertain outcomes (timed-out clicks, rerenders, stale element references) lead models to reason from wrong information or blindly repeat dangerous actions.
He open-sourced Chrome Controlla (Apache 2.0), a browser control layer for AI agents that:
- Tracks the exact Chrome tab and document the agent is working with
- Recovers stale references, scrolls off-screen elements back into view, and waits for unstable page regions to settle
- Never repeats actions with unknown results — instead reports honestly that it doesn't know if the first one happened
Benchmarks: ranked page snapshot 32ms; full snapshot of a 4,103-node page 103ms; verified a 3,973-character Google Docs write by checking the resulting document independently. Tested with Codex and Claude Code; the author argues trustworthy autonomy will come from harder-to-fool feedback loops, not just better models.
Related event: Chrome Controlla: open-source verification layer for browser agents(2 posts)→
More from coding & agent
- A scorecard for agent pilots: tasks, interruptions, missed decisions — krishnan · 2026-11-02
- Managing agents is babysitting: every ping arrives equally loud — krishnan · 2026-11-02
- A task you can't verify yet isn't done — agents need a 'deferred' task state — Key-Register-1969 · 2026-11-02
- A Redditor's ranked list of small models for running coding agents without a big GPU — transanethole · 2026-11-02
- Building revenue-team agents taught us context matters far more than model choice — hossam_elkhateeb11 · 2026-11-02
- "1% faster isn't enough, this should be 10x faster" becomes an AI-coding meme — DanielLockyer · 2026-11-02