Dejo Replays Website Tasks From Memory

Dejo’s compelling idea is turning browser automation from repeated rediscovery into a tested, stateful program—but its token savings are workload-specific, not universal.

• –Replays still consume roughly 89K tokens, largely from the agent’s own instructions and tool context rather than Dejo’s browser execution.
• –Every step is verified, fallback locators handle modest redesigns, and only the broken step needs relearning.
• –Task and identity checks reduce the risk of replaying an old workflow against the wrong account, subject, or action.
• –Dejo reports 19/19 correct answers, 78% fewer replay tokens than Playwright MCP, and one-call replays on 17 of 19 tasks.
• –The strongest fit is recurring workflows; one-off page lookups do not benefit much.