From meeting notes to follow-through

Microsoft's 18 September announcement covered agents in Teams, SharePoint and Viva Engage. The new collaborative agents entered public preview for Microsoft 365 Copilot users, while Facilitator for Teams meetings was generally available. Its examples connected notes and decisions with assigned tasks and project follow-up.

This changes what a project manager needs to assess. Accurate-looking minutes are insufficient if a task has the wrong owner, a decision loses its qualifications or an outdated document becomes the working reference. A review point is needed where information turns into action.

Agents for campaigns and delivery

Asana followed on 25 September with AI Teammates in beta. Its examples included drafting campaign briefs, checking material against brand guidelines and monitoring launch dependencies. Access, checkpoints and resource consumption were to be governed in the work platform. These were vendor-described uses, rather than independent evidence of realised benefits.

For a marketing department, the connection between brief, deliverable and approval matters. If an agent proposes content and then evaluates that same content, responsibility for final publication still needs to be explicit. Project records should also show which proposals a person accepted, so automatic updates do not conceal unresolved decisions.

Model progress does not establish operational reliability

Anthropic released Sonnet 4.5 on 29 September alongside the Claude Agent SDK and support for longer agent workflows. It highlighted its own coding and computer-use results. Those evaluations describe particular configurations; they do not establish how an agent will handle a Norwegian organisation's campaign plan or supplier follow-up.

TheAgentCompany research offers a different perspective. In the version published on 10 September, the strongest tested agent completed around 30 per cent of tasks in a simulated software company. That result concerned the study's models and environment, and did not test these three announcements. It nevertheless illustrates the distance between performing one step and completing an entire workflow.

Our assessment is that these offerings make bounded team trials more relevant. Measure completion, corrections and the time employees spend checking results. A campaign with named approvers and explicit deadlines provides a more verifiable trial than an open-ended instruction to improve marketing.

For discussion

Which project or campaign task has sufficiently clear ownership and review points for us to assess an agent's actual contribution?

Sources

Microsoft: Microsoft 365 Copilot – Enabling human-agent teams (September 18, 2025)

Asana: AI Teammates – Collaborative Agents That Deliver Results (September 25, 2025)

Anthropic: Introducing Claude Sonnet 4.5 (September 29, 2025)

Xu et al.: TheAgentCompany, version 3 (September 10, 2025)