October 2026 Summaries

No summary generated yet.

Modern CPUs use hyper-threading to run separate programs on two logical cores within one physical core, improving efficiency but creating potential transient-execution security risks such as MDS, where one workload could extract data from another. To protect virtual machines hosted for different exe.dev users, Linux core scheduling cookies restrict which processes may share a physical core. Rather than assigning every VM a unique cookie, which reduces hardware efficiency, exe.dev assigns a shared cookie to VMs within the same team, preventing VMs belonging to different teams or ungrouped users from running on the same core. This model assumes team members trust one another, while administrators of teams requiring stronger isolation can disable core sharing through the team settings command; unaffiliated VMs currently remain isolated from all other VMs by default.

exe.dev routes web requests to the appropriate virtual machine using the HTTP Host header, a mechanism central to modern web infrastructure such as load balancers, CDNs, and web servers. Users can connect custom domains to their VMs, with Let’s Encrypt certificates provisioned automatically, and can now also configure wildcard certificates such as *.example.com to direct all subdomains to one VM for custom routing. Wildcard domains can be added through the service’s lobby using the “domain add vm-name *.example.com” command, while DNS time-to-live settings may still delay activation; the post also corrects an earlier claim that Certificate Transparency caused such delays.

Software factory demonstrations often emphasize task management systems to coordinate multiple slow agents, but the author argues that high concurrency can disrupt human focus and merely conceal latency rather than solve it. They propose improving productivity with fewer agents through faster models, quicker testing, and interaction designs centered on human attention and information processing. At exe, experiments include using a fast, tool-free model called Luna 6 for real-time conversations while more capable models handle deeper work, though chat remains limiting for complex feedback. Shelley’s screen-recording feature lets users narrate while exploring a screen, then connects timestamped speech with visual context so agents can interpret richer, less structured feedback. For nonvisual engineering tasks, the author describes HTML artifacts that organize decisions, questions, examples, and diffs visually while recording user audio and in-page interactions through DOM tracking. Streaming partial feedback to agents during review could further hide latency by allowing responses and follow-up questions to appear before the user changes tasks, although the author says this approach has not yet matched the throughput of managing many agents.