2026-07-19

Daily AI Digest: OpenClaw reaches cloud workers as phone and DoorDash agents tackle errands

This is the GolemWorkers daily AI digest, tracking agents that can carry out real work while people keep decisive control. Today, OpenClaw reaches cloud workers and mobile devices, agentic phones coordinate travel and shopping, DoorDash opens ordering to agents, and new releases tighten delegation and research quality.

An operations professional supervises phone, laptop and delivery work from a sunlit office.
A field operations manager speaks to a phone beside connected cameras, a rugged laptop and a remote workstation.

OpenClaw sends agent sessions to cloud workers and mobile devices

OpenClaw 2026.7.2 beta.3 can place Control UI sessions on cloud workers, open Codex and Claude sessions in terminals on their owning hosts, and resume other coding sessions remotely. The beta also brings Automations to mobile, adds foreground Voice Wake on Android, and lets headless Linux nodes expose cameras, location and notifications.

Source: OpenClaw release notes
A traveler speaks to a smartphone while a ride, luggage and food pickup converge in one transit hall.

Agentic phones at Shanghai's AI show take on travel, shopping and messages

Nubia, Honor and StepFun showed phones designed around agents at the World Artificial Intelligence Conference. StepFun says its STEPX Neo can use partner services for travel bookings, purchases, local services, office work and video editing, while Nubia's NaviX Ultra is designed to work with participating apps instead of forcing access after earlier agent phones were blocked by major platforms.

Source: AFP via Tech Xplore
A team receives several unbranded lunch bags while one member checks the coordinated order on a laptop.

DoorDash opens food ordering to AI agents through a command-line beta

DoorDash's limited dd-cli beta lets agents search stores, find deals and complete checkout. Access is currently offered by waitlist to macOS developers in the United States and Canada, so they can build agents that prepare or place food and grocery orders without relying on the consumer app.

Source: TechCrunch
Two engineers inspect three separated compute lanes, including one clearly stopped task.

Cline prevents delegated coding agents from recruiting more workers

Cline CLI 3.0.45 removes the spawn tool from teammate agents so a delegated team cannot create more workers on its own, and failed teammate runs now report as failures instead of completed work. The release also cuts the global install from about 640 MB to 285 MB and retries once after refreshing expired sign-in credentials.

Source: Cline release notes
A researcher organizes many evidence folders across a vast archive while assistants retrieve more sources.

WANDR finds research agents still leave evidence sets incomplete

Perplexity released WANDR, a 500-task benchmark for research agents asked to build large collections and support every record with evidence. The strongest tested system reached 0.363 soft F1 and 0.133 hard F1 at high effort, showing that agents can find credible pieces while still leaving many professional research lists incomplete.

Source: Perplexity Research

Agent idea of the day

Let an agent plan the team lunch—then stop before checkout

Three colleagues review blank preference cards and meal choices before giving a final approval.

What this agent does

Collect lunch constraints, compare local options, prepare one complete group order and wait for a person to approve the exact basket before purchase.

Best for: Small teams with DoorDash dd-cli beta access that order recurring lunches and need to respect diets, budgets and delivery windows.

Give it

  • An opt-in headcount with dietary restrictions and allergy notes supplied by each person
  • The delivery address, arrival window, approved restaurants and maximum all-in budget
  • DoorDash dd-cli beta access plus one approved payment method that the agent cannot reveal or export

Tell it to

  1. Collect each participant's current preference privately and treat unanswered allergy questions as blockers.
  2. Search only approved restaurants that can meet the delivery window and every confirmed dietary constraint.
  3. Compare item prices, fees, taxes, tip and delivery timing against the all-in budget.
  4. Prepare one primary basket and one backup, listing substitutions, missing responses and the exact final total.
  5. Ask a designated person to approve the exact basket, total and address before checkout, then place one order and save the receipt.

Run it: Run two hours before a recurring team lunch or when a coordinator starts it manually for a one-off order.

You get

An approval-ready primary order, one backup, a constraint checklist and—only after approval—the final receipt and delivery status.

Keep a human in control

  • Never purchase until a person approves the exact basket, total, tip and delivery address.
  • Never infer allergies or dietary needs; send every ambiguity back to the participant or coordinator.
  • Keep individual dietary information private and share only the minimum needed to place the group order.
  • Never reveal, copy or export payment credentials, and never place a replacement order automatically.

Feasibility: DoorDash says the limited dd-cli beta can search stores, find deals and check out. Its launch demo showed an agent reading group context, inspecting menus, recovering from errors and calculating totals, which supports a bounded order-preparation workflow with a human purchase gate. Source: DoorDash dd-cli beta reported by TechCrunch →

GolemWorkers — OpenClaw ready to work: launch in seconds, no setup, online 24/7.