Agents get guardrails, new jobs and a backlog to clear

Security is gating enterprise agent adoption

Two security-focused posts point at the same constraint: agents behave differently from the users and services that existing controls were built for. Palo Alto Networks argues that autonomous, high-volume, cross-service calls from frontier AI systems break conventional network security assumptions, and it lays out design principles for treating agents as first-class network participants. Salesforce approaches the problem from the platform side, describing permissioning and guardrail changes meant to make Agentforce deployments safe by default so customers do not have to build heavy custom lockdowns. Salesforce frames security friction, not capability, as what has slowed enterprise adoption. That suggests secure defaults may matter as much as model quality in agent rollouts.

Agents move into docs, search and issues

Several launches extend agents beyond a chat window. WorkOS introduced Atlas, an AI coworker that plugs into existing workflows, which takes the company beyond its core authentication and infrastructure business. Cursor added Google Workspace plugins that let its assistant pull in and act on Docs and Sheets, stretching a coding tool toward general knowledge work. AWS added web search on Bedrock so foundation models can ground answers in live web content, which could replace some custom retrieval pipelines. It is also one more AWS change for teams to absorb, and a Techstrong piece asks whether customers can keep up with AWS's accelerating release pace without destabilizing production.

Cloudflare offers the most concrete example. It built a software factory that classifies, prioritizes and routes issues on the Astro project, aiming to drive the open GitHub issue count to zero. Maintainers of large open-source projects get a template for applying agents to backlog triage, not only to code suggestions.

Where human judgment still has to show up

As more work comes out of agents, some writers are asking what people should be doing around it. One post warns against becoming a meat proxy, someone who reviews, forwards or approves AI output without engaging with it, which undermines both the automation and the review. A long profile of Crosby describes a different approach. The law firm pairs lawyers from elite firms with engineers from top startups and embeds AI in its operating model for commercial contract work. It is a working case study for leaders planning AI-augmented teams.

A separate essay targets process rather than AI. It argues that debating epics and user stories distracts teams from outcomes, opportunities and measurable experiments.

Courtroom and stack pressure on AI vendors

Apple's motion for a preliminary injunction in its trade-secrets suit against OpenAI seeks faster relief than a full trial would provide. If granted, it could restrict OpenAI's product or hiring moves while the case continues, although initial reporting gives few details on which secrets are at issue. On a longer horizon, an essay on the endgame of vertical integration argues that advantage in AI is converging on companies that control the model, infrastructure and application layers together. It offers a frame for judging which vendors are likely to stay independent and which may get squeezed.