The week AI agents became coworkers

Something shifted in the first week of October. The framing of AI assistants moved, almost overnight, from "it answers your questions" to "it does your work." Here is what actually shipped, and why it matters even if you never buy any of it.

Microsoft: agents you pay by the task

Microsoft rebuilt Copilot around delegation. Cowork handles work you hand off, Code builds apps and dashboards from natural-language prompts, and Autopilot — an evolution of its Scout agent — runs in the background as a digital coworker with its own identity and configurable permissions. Word, Excel and PowerPoint are now embedded inside Copilot.

The important part is the billing. Chat and Office features stay under the fixed per-seat licence, but agentic work is billed by usage, in Copilot Credits. One setting deserves immediate attention: by default, new agents are added to your spending policy automatically. Cost control is now an admin task, not a footnote.

OpenAI: always-on "dots"

OpenAI introduced dots — always-on agents that pursue your goals across applications with limited supervision. They work through Slack and Teams, manage evolving projects, update proposals, analyse data and build software. OpenAI says they run on isolated cloud computers, and that sensitive actions (changing a password, permanently deleting data) require explicit consent. Weekly users of Codex and ChatGPT Work have passed 35 million.

Anthropic: cheaper, faster, and a talent bet

Anthropic released Sonnet 5.5, positioned as a faster, more economical model for coding and document work, and Haiku 5.5, its smallest and fastest model, aimed at high-volume small tasks and priced to make them cheap. It also launched the Claude Frontier Academy, committing $100 million to train 10,000 engineers — a bet that the scarce resource in AI is not models but people who can deploy them.

Google: data without copying it

Google shipped a preview that lets Gemini Enterprise query data directly in BigQuery, Cloud SQL and Spanner — without copying or ingesting it — using each user's own credentials over the open MCP standard. It is a quiet but significant shift: the model reaches the data instead of the data being duplicated into a store.

Why this matters to everyone

Whether or not you adopt any of these, three questions just became unavoidable for anyone building with AI:

• Cost. When work is billed by usage, a runaway agent is a runaway invoice. Budgets need caps and alerts, not good intentions.
• Permission. Agents now act, not just answer. What may they do without asking? The industry default is converging on "irreversible actions need consent" — make it explicit.
• Accountability. If an agent does something wrong, who is responsible? Regulators are already asking. The answer has to be a person, a policy, and a log.

The assistants became workers this week. The teams that do well with that will be the ones who decided, in advance, what the workers are allowed to touch.