OpenAI + Anthropic Are Pushing Agents Behind Your Logins. 3 Questions Every Project Team Must Answer.

OpenAI and Anthropic have both moved browser-capable agents closer to ordinary work, putting access, identity and approval controls at the centre of the conversation. Bill Gates has shifted the jobs debate towards policy, using construction to make the human stakes concrete. More than 100 companies are now calling for collective cyber defence as agent risks grow, while Anthropic’s reported $30T addressable market points to the scale of the investor story around AI work. In architecture, RIBA’s latest survey shows use has reached 74% of respondents, with formal governance and professional capability still catching up.

The practical difference between an AI assistant and an agent has narrowed again. OpenAI says ChatGPT Work can handle some tasks on signed-in websites through a separate cloud-browser session, with confirmation for consequential actions. Anthropic has introduced a browser within Claude Cowork, where a user can bring sign-ins into the agent’s separate browser on a site-by-site basis. The product designs and availability differ, yet both announcements move delegated browser activity closer to normal working routines.

For project teams, the relevant question is authority. An agent that reads a public planning page, accesses a document system, changes a record, or submits a form should not receive the same permissions. Define the task boundary, permitted sites and data, approval points, activity logs, and a fast way to withdraw access before connecting an agent to live systems. Vendor safeguards help, while each firm still owns the control design around its own project information and workflows.

Read the full story

Last week, Goldman Sachs put numbers on entry-level employment pressure. This week, Bill Gates has offered a policy response. His essay argues for payroll-equivalent taxes on automation and for areas of work society may choose to reserve for people. His example is a 55-year-old construction worker asked to move into elder care after a career in a trade. The proposal raises a useful question for employers: how will automation choices affect skills, identity and viable routes into expert roles?

Read More →

More than 100 organisations, including OpenAI, Anthropic, Google, Microsoft, CrowdStrike, Okta and Fortinet, have signed a letter warning that AI-enabled cyber attacks may become more widespread and sophisticated. The signatories call for collective action, improved observability, stronger access controls and greater support for defenders. For firms deploying agents, the practical lesson is close to home: establish scope, distinct credentials, containment, meaningful human review and a response plan before agents can act across business systems.

Read More →

Anthropic is reportedly preparing to tell potential IPO investors that its total addressable market exceeds $30 trillion. That figure describes a theoretical annual opportunity at full market capture. It does not describe revenue, valuation, buyer returns, or a forecast of sales. The investor narrative matters because it frames AI vendors as competing for the value of work itself. Buyers should keep the market-size headline separate from their own use-case evidence, commercial terms, adoption costs and governance requirements.

Read More →

RIBA’s third AI report finds that 74% of surveyed practices use AI on at least some current projects, up from 59% in 2025. The result is a strong signal from more than 1,100 self-selecting respondents, not a census of every practice. The open questions now sit around data, confidentiality, intellectual property, output review and early-career development. Adoption has moved quickly. The challenge for practices is to make their working controls, competence and accountability as visible as their tool choices.

Read More →

The 12 August paid release says seven homes are built at the 55-acre Salida scheme, where two RIC systems support construction. The figures describe a planned deployment, not completed output.

A PlanOps opinion article recounts an AI RAMS review that produced a confident verdict despite receiving no document. Its practical controls are evidence gates, valid references, visible evidence boundaries and human hold points.

Nima’s response to the MHCLG consultation argues for open standards, competence and information-management requirements. AI outputs inherit the quality of the information and process controls behind them.

Selected stories worth reading before you get back to delivery:

Snaptrude positions itself as an AI-native AEC workspace for site analysis, programming, massing, BIM and presentation. Its workflow starts with a project description, RFP or site, then lets a team refine the output and export to Rhino or Revit. The sensible first test is a low-risk brief with a clearly defined review stage. Check data settings, intellectual-property terms and output quality before adding live client information or relying on generated quantities.

Learn more →

OpenAI says California should add safeguards to its AI safety law, including incident monitoring during frontier-model training and stronger cybersecurity controls throughout the development lifecycle. The company had previously opposed the law. The policy shift is a useful marker of where agent governance is heading: monitoring, security and escalation processes are becoming part of the operating model, rather than optional paperwork after deployment.

Source →

The Dutch Data Protection Authority alleges that Uber deactivated some driver accounts through an automated process without adequate warning or human oversight. Uber disputes parts of the finding and says it will appeal. The lesson is relevant to workforce-management systems on projects and sites: any automated signal that affects someone’s work, access or livelihood needs a named accountable reviewer and a meaningful route to challenge a decision.

Source →

Charlie Hills has published a public guide based on 173 days of using Claude Code. It covers setup, planning, connectors, skills and routines, with an emphasis on practical use by non-developers. The guide is free to read, although using Claude Code requires the appropriate Anthropic plan. Read it for the core operating habits: begin with a bounded task, plan before changing files, and retain review points around anything consequential.

Source →

Sir Robert Buckland discusses the impact of AI on the justice system. [Podcast]

Sky News reports that surgeons at UCL Hospitals used live AI assistance during a brain-tumour operation on a British patient. The in-house tool, trained on hundreds of surgical videos, colour-coded tiny structures for surgeons to avoid. The clinical team remained responsible for the operation. It is a vivid example of AI contributing information at a critical moment while professional judgement remains central to the decision.

Source →

ChatGPT’s Library can surface connected Google Drive files and keep a Google Doc, Sheet or Slide open beside a conversation. Where supported and authorised, ChatGPT can update the source file; the initial experience excludes Shared Drives. Start with a non-sensitive monthly cost report or risk register. Ask for a comparison against the previous version, review the proposed changes, and keep the original record intact until a responsible project team member approves the update.

Source →

AI TechWorld runs this week alongside API World and CloudX. Its programme covers AI engineering, infrastructure, deployment, reliability, organisational adoption and governance. The event is relevant to this edition’s focus on browser-capable agents because the challenge is no longer limited to model capability. Teams also need dependable infrastructure, workable permissions and operating practices that hold up in live enterprise settings.

Details →

Till next time,

_{All content reflects our personal views and is not intended as professional advice or to represent any organisation.}