Last Week in AI: 7–13 September 2026

Small models for everyday work

Desert Ant Labs launched on 8 September with 18 on-device models, including transcription, audio cleanup and personal-data redaction, exposed through Swift, Kotlin and JavaScript SDKs. Twelve were labelled stable and six beta at launch.

local transcription or redaction could be worth testing in an internal tool. Measure missed names and technical terms on representative examples before relying on the output. Local processing alone does not validate the whole application's data handling.

AI joins network troubleshooting

Extreme Networks announced general availability of Agent ONE Coworker on 9 September for Extreme Platform ONE customers. It brings network context into recommendations and can proactively surface issues through its Nudge feature.

existing customers could compare its recommendations against a resolved incident. Check whether the evidence explains the fault and whether the proposed action would have helped. The vendor's announcement is evidence of availability, not our own verification of its performance.

A place for agents to run tools

Google made Computer Use and Shell sandboxes generally available in Gemini Enterprise Agent Platform on 9 September. The release includes network-boundary controls and customer-managed encryption keys.

teams building agents should explicitly decide where commands run and what those environments can reach. This provides managed infrastructure to evaluate; it does not make an agent's decisions correct or permissions appropriate automatically.

Another coding model to evaluate

Cognition released SWE-2 on 10 September, initially available in Devin Desktop and CLI, with other surfaces rolling out. Cognition reports competitive coding results at lower cost in its evaluations.

evaluate a small real maintenance task and count review time, corrections and successful completion. A vendor benchmark saving is not a promise that your team's total delivery cost falls by the same amount.

Closing question

Which would help your team most: local transcription, network investigation, a controlled place for agents to work, or help with routine code maintenance?