The editor has a night shift now.
VS Code 1.137 introduces Automations in preview: recurring agent tasks can run hourly, daily, or weekly, or on demand. The same release adds experimental voice interaction, lets a quick chat continue inside a workspace without losing its conversation, and brings issue and pull-request details into the Agents window.[1]
That is a bigger product move than another model picker. A coding agent is no longer only a person-shaped chat session. It can become scheduled infrastructure. Once the task wakes while nobody is staring at it, “it usually works” stops being a workflow.
A schedule is permission multiplied by time.
Measure the work, not the theater.
Bough takes a useful swing at the missing instrument panel. Its README says the local Go tool reads Claude Code and OpenAI Codex CLI session history, correlates that history with a repository’s Git log, and renders days, tasks, prompts, file changes, failures, and commits. It serves the view on loopback, offers text and JSON output, and says it performs no writes or network calls.[2]
The sharp bit is not the visualization. It is the separation between agent activity and surviving Git history. Bough explicitly warns those counts can disagree: an amend, rebase, or human-authored terminal commit changes what each record can prove. That is honest instrumentation. It also labels itself early, notes that task-grouping thresholds were fitted to one developer’s history, and keeps a struggle score off by default. Do not turn an interesting local lens into a universal productivity scoreboard.
Ownership is part of the stack.
Tailwind Labs announced that it is joining Shopify. Adam Wathan says Tailwind is installed more than 110 million times per week, will keep its open-source projects MIT-licensed, and will continue leading and maintaining them with Shopify’s support. Existing customers retain access to commercial products, while new sign-ups close as the team focuses on Tailwind CSS at Shopify.[3]
This is not directly an agent story. It is the other half of one. Software can be technically open and still depend on maintenance capacity, product pressure, and durable stewardship. Your automation stack has the same question at smaller scale: who owns the prompt, the tool permissions, the flaky fixture, the bill, and the cleanup when the original enthusiast moves on?
Put six fields on every unattended shift.
- Trigger: exact cadence or event, timezone, and the rule preventing duplicate starts.
- Scope: repository, branch, writable paths, network destinations, secrets, and spend ceiling.
- Owner: one person or team who receives failure and can disable the next run.
- Receipt: trigger time, versions, inputs, tool calls, diff, checks, failures, and exit reason.
- Review gate: what may publish automatically and what must wait for a named human.
- Stop drill: prove the schedule can be paused, credentials revoked, and partial work recovered.
If the system cannot print that card, it is not an automation yet. It is a recurring surprise.