As AI agents take on more real work, the teams that give them the most freedom first limit what the agents can reach. They also add a second check that does not depend on a person's quick glance, and design system teams and Atlassian apply the same idea to prototypes and plans.
David Cramer, a co-founder of Sentry, which makes error-monitoring software for developers, wrote about Junior, the company's internal AI agent that takes requests in Slack (13 Nov · 01). Cramer says the agent's name is on more code changes than any other single contributor at Sentry. In his example, he asked it in one Slack message to remove a failed experiment and gave no further detail. He glanced at its pull request, approved it himself and let the agent merge it, at a cost of \$1.32 and about five minutes of his time. He admits the change may have missed some references, which shows how thin the human check becomes once an agent handles routine changes.
Yuku Kotani, chief technology officer of Ubie, a Japanese health-technology company, described in Japanese the platform Ubie built so AI agents can work across the whole company (13 Nov · 02). The problem was sharing: an agent called in a shared Slack channel reads with the requester's access, so it could show management information to partners in that channel. Ubie split its agents into two kinds. Personal agents use one person's access and cannot be shared, while team agents run under service accounts limited to what everyone in the conversation may see. Because the agents' workspaces hold no credentials, they may make reversible changes without asking, and a nightly agent proposes improvements to the agents themselves.
Dan Scofield, writing on Shopify's engineering blog, described AI agents that attack Shopify's own live systems to find security flaws before outside attackers do (13 Nov · 03). Shopify, which runs online stores for merchants, says its bug bounty reports have risen almost 300% since 2024 as AI models made finding flaws cheaper. A second AI agent, running on a different AI model, reads everything the attacking agent does and stops it if it breaks the agreed rules. If that second agent fails, the run ends, and Shopify reports tens of thousands of runs with no outages caused by the testing.
Five Pinterest engineers, data scientists and managers, including software engineer Skyler Zheng, described how they made the company's data-analysis AI agent follow its required steps (13 Nov · 04). Rewriting its instructions by hand was trial and error, and one sensible rule made it refuse a question it could answer. The team built paired tests from real chats, one set where a rule applies and one where it does not. An automated loop now keeps only changes that fix failures without breaking anything else. Because 10.3% of cases changed result between identical runs, they judge changes case by case, not by the total score.
Christine Vallaure, a designer who teaches design and code through her learning platform moonlearning, described how she moves a prototype between code and Figma with an AI agent (13 Nov · 05). Her one condition is that every component has the same name and properties on both sides, so the agent looks parts up instead of guessing. She asks the agent to leave a note listing what it built by hand and where the two sides differ, and she reads that note before the screens. The round trip caught a 21-pixel spacing fault in the code that no automated check had found.
Kelsey McKeon of Retool, a company that sells software for building internal tools and AI agents, reported its survey of 300 technology leaders (13 Nov · 06). Half of the leaders with agents in production named agents making decisions with no clear owner as a top concern. Thirteen percent said an agent had taken an unintended action in a production system in the past year. Retool recommends a named owner and a defined scope for every agent before launch, and its figures come from a vendor with an interest in the answer.
Christine Vallaure (13 Nov · 07) also wrote a guide in UX Collective on why AI builders connected to a design system still hand developers a lookalike. When she tested Claude Design on her own code, every button property and token came back right, but its helpers, styles and types were replaced. Her remedy is a list, generated from the code, of every component, property and allowed value. The agent must query it through an MCP server, a standard connection that lets an agent ask a tool what exists. Prototypes stay on a design branch that is never merged, and accepted ideas are rebuilt and checked before they reach the product.
Zalando's design system team built its own AI prototyping tool, the ZDS Capsule, after Figma Make kept producing screens that were off brand. Dana Baddad and Gonzalo Vásquez Correa (13 Nov · 08) described it at a meetup in Berlin, in notes published by the newsletter Into Design Systems. The team rewrote its guidelines as small markdown files and made its skills run scripts at critical steps, because the agents' choices varied from run to run. More than 20 designers have contributed, but an engineer found the AI-written motion specs for one feature hard to follow. The team judged that implementation could take two or three times as long.
Tim Gremore (13 Nov · 09), development lead at the agency Headway, wrote about what a design system costs when AI agents read it. A DESIGN.md file, the usual written rules for an agent, is read in full on every turn. In a test Atlassian published, an 80 KB file used almost twice the tokens of a larger MCP server fetched in slices. Headway runs its own MCP servers as local packages that answer narrow questions, such as which properties a component takes.
Avani Prabhakar (13 Nov · 10), Atlassian's Chief People and AI Enablement Officer, described how the company is reworking product development for nearly 10,000 research and development staff. Its own survey found that 26% of developers now say AI saves them more than ten hours a week, down from 68% a year earlier. Agents now prepare fixes for 52% of security vulnerabilities, so engineers start from a proposed fix and decide whether it is safe to ship. Decisions are kept in shared specifications that become the tasks agents carry out. More than 50 product managers and designers shipped over 100 engineering-approved changes in three months.
Jennifer Riggins (13 Nov · 11), writing for LeadDev, asked developer-experience researchers how engineering organisations should measure AI agents, with no DORA survey of software delivery this year. Their answers stress developers' wellbeing and the cost of checking agents' work, which a McKinsey report put at 60% of agentic AI spending. Although 87% of organisations claim AI governance policies, fewer than a quarter have put controls in place. Charlotte Fleming of Octopus Deploy argues that the rules must run as gates in the deployment pipeline.
Ubie and Shopify let their agents act with little supervision only after limiting what the agents could reach and adding a check that does not tire. Pinterest built the same kind of check for an agent's behaviour. Vallaure, Zalando and Headway built theirs for design systems, as a list of real components or a script the agent must run. Atlassian applies the idea to the whole process, keeping decisions in specifications the agents work from and people at the points that need judgement. The inference here is that where those limits are missing, as in Cramer's account, one person's glance is the only check. Retool's survey and the researchers LeadDev asked suggest many organisations are in that position.
Most of these accounts come from the teams that built the systems, and only Pinterest reports careful measurements. Atlassian's figures come from its own pilots, and Headway's cost comparison rests on one Atlassian test of one screen. Zalando's experience also shows that a faster prototype can make implementation slower. Sentry's light review may be the more honest picture of most teams, and for low-risk changes it may be good enough.
David Cramer, a co-founder of Sentry, says the company's Slack-based AI agent has more code changes to its name than any person, often merged after a brief look.
Ubie, a Japanese health-technology company, lets AI agents work without approvals by limiting each shared agent's access to what everyone in the conversation may see.
Shopify built AI agents that attack its own live systems to find security flaws, with a second AI agent on a different AI model ready to stop them.
Pinterest engineers stopped hand-editing their data agent's instructions and built an automated loop that keeps only changes passing paired tests drawn from real chats.
Designer Christine Vallaure moves prototypes between code and Figma with an AI agent by naming components identically on both sides and asking for a note of anything invented.
Retool's survey of 300 technology leaders found half worry about AI agents making decisions with no clear owner, and recommends naming an owner and scope before launch.
Designer Christine Vallaure found that AI builders connected to a design system rebuild lookalike components, and says agents need a generated component list they must query.
Zalando's design system team built its own AI prototyping tool, the ZDS Capsule, after Figma Make stayed off brand, then learned its prototypes could slow implementation.
Tim Gremore of the agency Headway says a design system's AI costs depend on how often its context loads; in an Atlassian test, a smaller file cost more.
Atlassian is redesigning how its nearly 10,000 research and development staff build products with AI agents, aiming people at decisions and keeping shared specifications the agents work from.
Journalist Jennifer Riggins asked developer-experience researchers how to measure AI agents' effect on engineering teams; they point to wellbeing, enforced accountability and the cost of checking agents' work.










