Anthropic AI Updates: July 23, 2026

1. Anthropic Publishes a Research Agenda for Its Economic Futures Research Fund

Anthropic. Anthropic published a research agenda for its Economic Futures Research Fund, laying out five priority areas: workplace AI integration, workforce transitions, modernized income support, worker equity stakes in AI-driven gains, and evidence to guide public investment. The agenda invites large-scale proposals from universities, research institutes, and nonprofits, with individual projects ranging from roughly $5 million to $30 million. It reflects the company’s effort to direct funded research toward measuring and preparing for AI’s economic effects. Source

2. Claude Gains a Connector for Querying the Anthropic Economic Index

Anthropic. Anthropic released a connector that lets users query the Anthropic Economic Index directly inside Claude on claude.ai, with no installation required. People can ask questions such as which occupations use AI the most and receive answers grounded in the Index’s underlying usage data. The tool is aimed at making the Index’s findings on AI adoption accessible to researchers, journalists, policymakers, and general users through conversation. Source

3. Anthropic Donates Another $20 Million to Public First Action

Anthropic. Anthropic said it is donating another $20 million to Public First Action, bringing its total support for the group to $40 million. Public First Action educates the public about AI and works across party lines with Republicans, Democrats, and Independents on AI policy. Anthropic framed the donation around the need for government safeguards as model capabilities continue to advance. Source

4. Anthropic Details How to Build Verification Loops in Claude Code With Skills

Anthropic. A new Claude Code post walks through turning manual verification steps into reusable skills, markdown files placed in .claude/skills/ that instruct Claude to check and fix its own work. It covers built-in mechanisms such as the /verify skill, toolchain integration, and automated code review, then outlines four patterns for applying custom skills: standalone, embedded, chained, and PR-wide. The guidance positions repeatable verification as a way to keep agent-written code reliable without constant human correction. Source

5. Outtake Builds an Autonomous Cyber Investigator on Claude

Anthropic. In a Claude case study, cybersecurity startup Outtake described its Recon Agent, an autonomous investigator built on Claude that maps the networks behind phishing pages and other impersonation attacks. The agent runs for a median of about 16 minutes and occasionally up to two hours per investigation, prototyped in Claude Code before moving to production on the Agent SDK. Outtake’s takeaways include treating prompts as flexible suggestions rather than hard constraints, giving agents open-ended tools such as filesystem access, and building evals from the start. Source