AI agents enter spreadsheets, metrics and legacy apps as token costs tighten
Writer rebuilt its harness to constrain token spending, Google added agent-driven interfaces to Sheets and BigQuery, and AWS brought legacy web applications into the same workflow through a managed browser.
Artificial Intelligence··Midday
A rebuilt harness to hold down token spending
Writer released Palmyra X6 on Thursday alongside a rebuilt agentic harness. Palmyra X6 is a post-training variation on Z.ai's open-source GLM-5.2, and the harness work targets multi-step tasks that burn tokens without finishing. The company puts the average cost reduction from the harness changes alone at 40 per cent, and says basic tasks can run up to 50 per cent cheaper. Writer's researchers write that harness efficiency changes proved more reliable than model selection for cutting cost, and that the harness multiplies efficiency across every model an organisation runs. Chief executive May Habib said enterprises are sick of chasing the next benchmark and want flattening cost, and that chief information officers are giving up on the labs. The model reached Writer's clients on Thursday. Every figure here is the company's own and no independent evaluation has been published. When token spending becomes as visible a constraint as the model itself, a new model release has to be read together with cost control.[1]
Layout on the sheet, graph path behind the metric
Google added Sheets canvas, a tab inside Google Sheets where a request to Gemini builds an interactive layout over data already in the sheet. It opened first to Google AI Pro and Ultra subscribers, with Workspace Business and Enterprise Standard and Plus plans starting the same day. Canvas and sheet stay in sync both ways, so editing the data changes the view, and the view can be shared with the same permissions as any other spreadsheet. Google's examples include a study tracker, a fantasy football dashboard and a wedding seating chart; users ask Gemini to change layout, design or functionality. Subscribers to the Google AI Pro for Education add-on get it as well. In the same window Google Cloud put governed measures into BigQuery Graph in preview, so an agent can aggregate a metric along property-graph paths instead of reading a flat table. The engine resolves the graph path first and evaluates the metric afterwards. In one retailer example, an agent can report a 12 per cent fall in winter jacket sales in Seattle from a flat table but cannot follow the path from orders to distribution centres to suppliers held up by storms, and therefore proposes a 15 per cent markdown that erodes margin. Measures are declared with SUM or AVG inside the property graph definition and queried through GRAPH_EXPAND, mapping existing tables in place without a separate load. BigQuery Studio adds a drag-and-drop graph modeller and a natural-language query path; graphs can also be defined in LookML under version control.[2], [3]
A managed browser for legacy web applications
AWS published a reference build in which an agent works inside Amazon Bedrock AgentCore Browser Tool, a managed cloud browser, to operate web applications that expose no interface of their own. Sessions run isolated, and the recording of each one goes to Amazon S3. The agent connects to Chrome over the Chrome DevTools Protocol with Playwright and reads the screen through a vision-capable model on Amazon Bedrock, so it does not depend on the application's technology stack. Browser profiles carry authentication state between sessions, proxy settings route traffic through corporate infrastructure, and the workflow can pause for an operator to confirm a step. The reference implementation deploys as one Terraform stack with a React interface, an NGINX proxy on Fargate, the AgentCore runtime and isolated Chrome instances, with sign-in through Amazon Cognito and OIDC providers. Writer's harness for token spending, Google's sheet and metric surfaces, and AWS's browser path into legacy applications widen the layers an agent can enter while making cost and control part of the same workflow.[1], [2], [3], [4]
