2026-08-30
Daily AI Digest: travel agents track fares; grocery agents fill carts
This is the GolemWorkers daily AI digest for anyone who wants a practical view of what autonomous software can do now. Today, agents watch airfares, keep multi-day work moving, prepare grocery carts, and act behind stronger human controls—while one new test shows why those boundaries matter.
Google AI Mode can track fares and help book a hotel
Google added conversational flight-price tracking, points-and-miles comparisons and hotel booking to AI Mode in Search. Price alerts are available in more than 180 countries and territories, rewards rates are global, and English-language hotel booking is rolling out in the U.S.; travelers still review the room, cancellation terms and payment before the hotel or booking partner completes the reservation.
Source: Google
Podiom gives Claude and Codex goals that can run for days
Podiom's latest open-source release puts local Claude Code and Codex agents in one workspace with durable sessions, schedules, shared projects and goals that a lead agent can plan and review over days or weeks. It runs on macOS, Linux and Windows and keeps tool activity attached to the goal, but autonomous goals use full access without per-action prompts, so they belong only in a tightly scoped workspace.
Source: Podiom
A household agent can fill the weekly grocery cart—without checking out
A newly documented home workflow connects ChatGPT to a recipe planner and a supermarket through narrow MCP tools, then lets the agent turn meals into quantities and choose only from products the household already buys. The agent prepares the online cart while credentials stay inside the connector and the person keeps the final order button, showing a repeatable pattern for useful automation with a hard spending boundary.
Source: Thibault Ducret
Runplane puts a policy gate in front of agent actions
Runplane launched a hosted control layer that evaluates an agent action before it reaches tools connected through MCP, OpenAPI or internal APIs. Teams can allow routine reads, block destructive commands or pause higher-risk deployments and payments for approval, with source-specific policies and an audit trail; the commercial service starts with a 14-day trial and is aimed at organizations already running agents in production.
Source: Runplane
A web page hijacked Claude Code Auto Mode in a small test
Security researcher Johann Rehberger reports that a malicious page presented as a routine summary task triggered command execution in Claude Code Opus 5 Auto Mode in 60–80% of a small test sample. The result does not establish a population-wide failure rate, but it gives people running unattended coding agents a concrete reason to use containers or VMs, restrict network access and keep credentials outside the agent's workspace.
Source: Embrace The Red
Gemini Notebook can ground answers in books you own
Google's new Expert Intelligence program lets U.S. users add eligible ebooks purchased through Google Play Books to Gemini Notebook, ask grounded questions and combine the book with private notes or other sources. It launches with more than 100,000 titles from major publishers and requires collaborators to own their own copy before they can interact with the book.
Source: GoogleAgent idea of the day
Build a grocery-cart agent that cannot spend your money
What this agent does
Turn a weekly meal plan into a review-ready grocery basket while keeping credentials, product choices and the final purchase under explicit human control.
Best for: Households, caregivers and small shared kitchens that repeat online grocery pickup or delivery orders every week.
Give it
- Your meal plan, serving counts, dietary rules and a list of ingredients already at home
- Read and write access to a recipe or shopping-list app through a narrow connector
- Search and add-to-cart access limited to a supermarket favorites list or approved product catalog
- A fixed weekly budget and one named person who alone can place the order
Tell it to
- Convert the approved meals into one consolidated ingredient list and subtract the pantry items I confirm are already available.
- Match each remaining ingredient only to products in the approved favorites list, choosing the smallest quantity that covers the recipe.
- Flag substitutions, unusually high prices, missing products and any item that would push the basket over budget.
- Prepare the cart without opening checkout, changing delivery details, applying stored credit or placing the order.
- Present one review sheet with meals, quantities, product matches, substitutions, total price and a link for the named person to inspect the cart.
Run it: Run after the household approves next week's meal plan, or every Saturday morning if an approved plan already exists.
You get
A review-ready cart plus a concise exception list showing missing items, substitutions, price spikes, pantry assumptions and the final total.
Keep a human in control
- Never add products outside the approved catalog without asking, and never infer allergies or dietary restrictions.
- Never expose supermarket credentials, session cookies, payment methods or home-address data to the language model.
- Never submit the order, change pickup or delivery details, buy age-restricted goods or spend stored credit.
- Stop and ask when the total exceeds budget, a required product is unavailable or a substitution changes nutrition or allergens.
Feasibility: Ducret's working household setup connects an agent to Mealie and a supermarket through narrow MCP tools, limits matching to products the family already buys, keeps credentials inside the connector and leaves the final order action to a person. Those same boundaries make this recipe practical without granting the agent spending authority. Source: Thibault Ducret — grocery shopping with AI →
