BenchFlow
The agent that runs on the phone in your pocket

An end-to-end recipe for a small mobile agent that actually converges.

Built on env0 — synthetic Google Workspace environments, a synthesizer–judge data game, an SFT→RFT pipeline, and a quantized Gemma that runs on-device. The whole loop, open.

0 → 2
of 60 tasks, after fine-tune
≈4B
params · runs on a phone
4
mock envs · gmail/gcal/gdoc/gdrive

What we’re shipping

Four pieces, one pipeline.

What it can do

General-purpose mobile use — verified, end to end.

Sample trajectoryreward 1.0 ✓
  1. User

    “Find the dinner Maya suggested in email and put it on my calendar.”

  2. Agent

    Search the inbox for Maya's recent threads about dinner.

  3. Toolgmail.search

    q="from:maya dinner" → 3 threads, 1 with a proposed time

  4. Agent

    Extract the restaurant and the proposed Friday 7:30pm slot.

  5. Toolgcal.list_events

    Friday 19:00–21:00 → no conflicts

  6. Toolgcal.create_event

    “Dinner with Maya” · Fri 19:30 · Tartine

  7. Verify

    event exists, correct title + time, no double-book → reward 1.0

Also handles

  • Search emails & calendar
  • Find restaurant / flight / event info
  • Send mail & manage invites
  • Take notes & prep for meetings
  • Drive: tally calories from the last 24h
  • Research & summarize

Every task runs against a stateful mock of the real API and is graded by a deterministic verifier — so a passing trajectory means the work actually got done, not that it looked plausible.