The Agent Finished. Who Approved the Result?

Alistair Barr’s workplace test of OpenAI’s Dots produced something more useful than a theatrical demo: a newsletter draft his editor judged largely correct, including photos, a chart and links. It also produced a website that the agent could not make public because the corporate workspace prohibited external sharing. The same test delivered both useful execution and an enforced limit. Apparently, the future still needs to speak to the administrator. Business Insider

The newsletter result came after a specific sequence. Barr obtained workplace approval, connected applications, supplied a training guide and an old newsletter, entered his password and participated in Okta two-factor authentication. The agent then populated the publishing tool and returned a test draft for inspection. This was a workflow with instructions, authorized access and an editor at the receiving end—not evidence that every corporate task can now be handed to an owl avatar. Business Insider

The editor’s informal assessment is encouraging precisely because the output was inspectable. Moving content into a complicated form is work worth removing from an editor’s day; checking the resulting draft remains a separate job. Meanwhile, the blocked website demonstrates that generating an artifact and obtaining permission to distribute it are different operations. For CTOs, that distinction makes this test useful: evaluate whether the agent performs the particular task, then decide which consequences its access should permit. A successful draft should earn another evaluation, not a master key.

A purchase can complete successfully while the request remains poorly specified. Austin Welch asked Instinct to order chicken wings and drumsticks for dinner with his partner, omitted the quantity and received twenty pounds of wings. His theory about the agent drawing on previous purchases is speculation; the concrete result is sufficient. The checkout worked. The dinner specification did not. This is an unusually edible illustration of why task completion is an inadequate acceptance criterion. Business Insider

Other shoppers described ways to retain control over payment. Loni Stark said her Muse agent obtained permission before buying ballet shoes. Amando Komoda said he approves purchases in advance and uses a single-use credit card; he also often buys items himself, partly because Instinct cannot always complete checkout Captchas. These are accounts from individual users, not measured reliability results. They nevertheless show delegation being divided into product discovery, transaction preparation and permission to spend. Business Insider

That division complements the Dots workspace restriction: both place boundaries around an action, though one concerns payment and the other external sharing. Neither boundary establishes that the proposed action is correct. A permission check cannot resolve an unspecified grocery quantity unless the approval process actually makes that quantity visible and reviewable. Dots’ blocked publication and the shoppers’ payment practices therefore suggest two separate questions for deployment: may the agent do this, and has it prepared the right thing to do? Combining them into a single success flag is how a system can pass its workflow and fail dinner. Business Insider

Vesta’s $30 million funding round takes the delegation question somewhere less forgiving than a publishing form or grocery basket. Three customer lenders joined the investment, and reported uses extend from document review and data validation to underwriting decisions at some lenders. That range matters: assistance with supporting documents and participation in a consequential decision demand different scrutiny, even when both appear under the same agent branding. TNW

CEO Mike Yu credits Claude Sonnet 4.5 with following configured instructions over the long stretches mortgage work requires. He also says lenders retain underwriting accountability and that actions and their reasoning are logged for audit. Those are vendor claims, not an independent evaluation of decision quality. Still, the retained responsibility makes the operational question concrete: can the lender establish that the delegated work supports a defensible decision? A record of the process helps make it inspectable; its existence alone does not establish that the outcome was sound. An audit trail is not a passing grade written in smaller type. TNW

The report also points to GDPR Article 22’s existing restriction on solely automated decisions with legal or similarly significant effects. It does not establish that Vesta’s customers make wholly automated decisions or that Vesta has violated that rule. The relevant implication is narrower: the consequences and decision process constrain acceptable delegation. TNW Barr’s newsletter test ended with an editor inspecting an intermediate artifact; that is a useful model for that workflow, not a universal approval architecture. Business Insider Before expanding access, CTOs should specify the permitted steps, the outputs that require checking and the person who owns the final decision. The agent can finish its assignment while the organization’s responsibility remains very much in progress.