2026-07-31

Daily AI Digest: Robot agents coordinate; AWS shifts course

This is the GolemWorkers Daily AI Digest for July 31, covering five practical shifts in how agents coordinate robots, move between cloud platforms, organize coding work and connect to tools. A separate field guide shows how retailers can turn product knowledge into an after-hours adviser without handing over final decisions.

An operator reviews a product sample and robotic component while a service robot works in a modern operations studio.
Two mobile robotic arms coordinate an object-sorting task while a researcher supervises from outside the safety zone.

Google’s new robotics model can coordinate multiple robots

Google released Gemini Robotics ER 2 in public preview through two Gemini API endpoints for robotics teams. The models accept text, image, video and audio, support multi-step tool use and multi-robot coordination, and can pause through blocking function calls before a physical action; a streaming version adds low-latency audio and video for real-time robot agents.

Source: Google AI for Developers
Two engineers verify a controlled migration between separate data-center infrastructure zones.

AWS closes Bedrock Agents Classic to new customers

As of July 30, new AWS customers can no longer create Bedrock Agents Classic, while existing agents keep running. AWS recommends AgentCore for new work and migrations; its managed harness handles compute, memory, identity, observability and human-in-the-loop pauses, and an AWS agent skill can inspect a Classic setup and prepare a migration plan before deployment.

Source: Amazon Web Services
A software professional organizes parallel project folders, compact work computers and approved connector modules at one desk.

Codex adds reusable agent plugins and better thread control

OpenAI’s Codex 0.146.0 release lets users name and pin sessions, keep side conversations open while switching threads, and fork histories without losing the original work. It also adds Agent Plugins manifests, workspace plugin publishing, marketplaces for Amazon Bedrock and Claude Code, secure discovery of executor-provided skills and remote Code Mode connections over WebSocket.

Source: OpenAI Codex
A systems architect checks an approval card before connecting standardized hardware interfaces in an integration lab.

MCP’s new stable standard gives agents cleaner handoffs

The Model Context Protocol’s 2026-07-28 revision is now stable for products that connect agents to tools and data. It makes requests stateless, adds server discovery, moves long-running work into an official tasks extension, and introduces an explicit input-required result so an agent can pause, request a human decision and retry with the answer instead of relying on a server-initiated callback.

Source: Model Context Protocol
A founder and analyst review a large stack of routine work beside a compact compute system and an open budget ledger.

OpenAI cuts GPT-5.6 Luna’s price by 80%

OpenAI reduced GPT-5.6 Luna pricing by 80% and GPT-5.6 Terra pricing by 20% on July 30. Luna is the family’s lowest-cost model and Terra is positioned for everyday work, so teams using them for high-volume research, support, document and tool-driven tasks can run the same workloads at a lower model cost without changing model families.

Source: OpenAI

Agent idea of the day

Build an after-hours product adviser with a human handoff

A shopper speaks to an after-hours product adviser beside small appliances while a store owner reviews the proposed shortlist.

What this agent does

Turn an open-ended product question into a short, source-backed comparison that a shopper can use after hours and a store employee can review before any commitment.

Best for: Independent retailers and service businesses with a reliable catalog, repeat customer questions and limited evening support coverage.

Give it

  • An approved product catalog with current prices, dimensions, availability, warranties and return rules
  • A short interview guide for budget, space, must-have features and deal-breakers
  • Clear escalation rules for uncertain information, regulated advice, discounts, payments and final recommendations

Tell it to

  1. Ask one question at a time until you understand the shopper’s budget, constraints and priorities.
  2. Search only the approved catalog and quote the product record behind every factual comparison.
  3. Prepare no more than three options and explain the tradeoff that separates each one.
  4. Flag missing or stale price, stock, warranty or compatibility data instead of guessing.
  5. Send the shortlist and conversation summary to a staff member whenever the shopper wants to buy, negotiate, finance or rely on a consequential recommendation.

Run it: Run when a shopper starts a voice or chat conversation outside staffed hours, or when the live queue exceeds the team’s response target.

You get

A three-option shortlist with source links, constraints matched, unresolved questions and a clear human-review handoff.

Keep a human in control

  • Never take payment, change a price, promise stock or accept contract terms.
  • Do not infer sensitive traits or use them to rank products.
  • Keep every factual claim tied to a current approved catalog record.
  • Require a person to approve discounts, financing, regulated advice and the final purchase handoff.

Feasibility: OpenAI reports that avatarin and Yamada Holdings used GPT-Realtime to run a multilingual voice-and-text shopping agent grounded in product information and trained on store-associate knowledge. About 30,000 people used it during a two-week online-store campaign, and the agent asked follow-up questions to narrow needs while guardrails kept the conversation focused on shopping. Source: OpenAI →

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