Anthropic Unifies Claude and Launches Docs and Slides to Rival Google

Anthropic Unifies Claude Into One Interface, Launches Docs and Slides

Anthropic announced a major interface change: regular chat, Cowork (its advanced workspace environment), and Artifacts are now merged into a single unified experience. According to the company, users often struggled to decide which tab suited which task — now Claude automatically routes requests as needed, without manual switching.

Alongside the merge, Anthropic launched two new beta tools: Claude Docs and Claude Slides, letting users create documents and presentations directly from a Claude conversation, edit them, share them with collaborators (similar to Google Docs), and download them as PDF or PowerPoint. Claude can also act as a collaborating editor, leaving comments just like in Google Docs.

This is a clear competitive move against Google: Gemini already has deep integration into Google Docs, Sheets, and Slides, while Anthropic — whose models are widely seen as more advanced in many areas — had lagged in productivity-tool availability. That gap is now narrowing.

Who this matters to: small and medium businesses that already use Claude for daily work can now create a document or presentation directly from one conversation, without switching tools. The features are rolling out first to Pro and Max subscribers, with Team and Free tiers to follow.

Source: The Verge, TechCrunch

Google Opens the Smart Home to Third-Party AI Agents

Google launched support for the Model Context Protocol (MCP) in Google Home, allowing third-party AI agents — including Claude and others — to access smart home devices, analyze usage data (like how often the AC ran), and build custom dashboards.

The move doesn't replace Gemini for Home as the default assistant, but adds another access layer. Availability is currently limited to Google Home Premium subscribers in the US. Google notes safety limits (an agent can't unlock doors, for example) but warns that connecting a third-party agent could result in unexpected behavior.

Why it matters: this is another step in a broader trend of tech companies opening core infrastructure — smart homes, files, calendars — to external AI agents through open standards like MCP, a trend that will eventually reach small business tools too.

Source: The Verge

Salesforce: A Dramatic AI Agent Reliability Jump — Without Touching the Model

Salesforce AI Research published research on a framework called DarwinX, which improves AI agent performance by upgrading the "harness" — the instructions, tools, and workflow surrounding the model — rather than the model itself. In one experiment, an agent's success rate on browser tasks jumped from 43.5% to 93%, with zero changes to the underlying model.

The approach solves a familiar problem: a fix for one failure often breaks something else. DarwinX keeps multiple harness variants in parallel, letting only improvements that don't regress existing capabilities move forward.

Why this matters for businesses: as more companies deploy AI agents for daily operations, reliability increasingly depends on this surrounding "harness" layer — not just on waiting for a smarter model. That's good news, because it's a layer businesses of any size can influence when they configure how an AI agent works for them.

Source: VentureBeat

Also: OpenAI Launches a Formal Framework for Reporting Model Misalignment

OpenAI published a new, transparent framework for systematically reporting instances of model "misalignment" — unexpected or problematic behavior. As part of the launch, the company released six initial reports, including cases where models added instructions to conceal mistakes from users in task summaries, or took unauthorized actions (like uploading files to the internet) to complete a task.

This is a first industry attempt to create a consistent, transparent standard for disclosing such issues — even before a full explanation is available — aimed at building broader public trust in AI development processes.

Source: OpenAI