Skip to content
Article Every Oct 2026

Every: my first week with an AI coworker

Every has built an AI agent — it calls it the Every Agent — that works in the team’s Slack alongside human staff. When Grant Harvey joined as a staff writer in October 2026, it was his first time working in an office where an AI agent functioned as a standing team member. His account of that first week, published in Every’s Context Window on October 29, 2026, is a practical field report rather than a product pitch: what the agent was actually useful for, where it fell short, and what that suggests about how to split the work.

What the agent did well

Harvey found the agent reliable for retrieval and context tasks: locating documents, surfacing past discussions, and finding relevant prior writing. For a writer getting up to speed at a new company, that kind of quick-lookup capability is genuinely useful. The agent could move through the team’s internal knowledge faster than a new hire scanning through Slack archives manually.

Where it went wrong

The clearest failure came when the agent reported outdated company information as current. The article does not dramatize this — it treats it as one data point in a workflow assessment — but the implication is straightforward: the agent does not know what it does not know, and verifying its output on anything that depends on being current remains the writer’s job.

The practical division

Harvey arrives at a three-way split that the article presents as his main takeaway: use the agent for context-gathering (documents, prior work, company knowledge), use a dedicated chat model for drafting, and ask people for judgment calls and sign-off. He names GPT-6.1 Terra and Gemini 4 Argon with different assessments: the former as a capable iterative writer that needs significant human direction, the latter as more useful as an editor than a ghostwriter.

This is not a novel framework, but Harvey’s account gives it specific texture — where each type of tool performed as expected and where it did not.

A note on AI-influenced decisions

The piece cites an Anthropic study finding that analysts who used Claude adopted its preferred approach 86% of the time and rarely credited the AI when explaining their reasoning. Harvey takes this as a reason to track your own decision-making consciously: not just the output, but the reasoning behind the calls you make. For writers whose work involves judgment — argument, editorial stance, source selection — this is a genuinely useful point.

Who it’s for

Useful for writers and editors evaluating AI agent workflows in a team context, for newsletter or content teams considering adding an always-on AI to their tools, and for anyone interested in what day-to-day collaboration with an AI agent looks like in a writing-focused organization.