Three announcements landed within a day of each other, and they point at the same problem from different directions. GitHub made Copilot part of GitHub Enterprise and moved agents into repositories as scheduled jobs. ServiceNow gave non-coders a way to build agents on top of enterprise systems. Replit shipped Cloud Agents that let a support lead open a pull request against your application.
The models are the least interesting part of any of this. What matters is who is allowed to create these agents, what they can change, and who reviews the output.
What GitHub actually shipped
Copilot is now included with GitHub Enterprise. All Copilot features, no separate subscription, paid through AI credits. It's available through GitHub sales now, with self-serve "coming soon," and existing Copilot Business and Enterprise customers can stay on their current plans or move later. That last clause is doing a lot of work: nobody has to renegotiate this week.
Copilot automations are the piece platform teams should read twice. These are AI workflows that run in the cloud inside a repository, triggered by hand, by an event, or on a schedule. GitHub's own examples are issue triage and code review. Automations can suggest changes to themselves based on past results, which a person then approves.
Three other changes came with it. Copilot code review Max adds a deeper review tier for high-severity issues alongside the existing Lite and Balanced tiers. Copilot Fusion is a routing layer that picks a model per part of a task; I Programmer reports it as generally available, and GitHub says its top setting scored within two points of the strongest model it compared against on a public coding agent benchmark, at half the cost and 40% less time per task. Treat that number the way you'd treat any vendor benchmark on a public task set, because your monorepo is not a public task set. Copilot projects, available in the Copilot app since 28 October, runs agent work across several repositories as parallel sessions with one place to track progress.
The DevOps Society editorial on the launch has a checklist worth stealing:
- Get credit prices in writing and compare them against your current seat costs before agreeing to move.
- Set budgets before enabling features. Code Security AI credits can be budgeted at SKU level, so do that first, then turn on AI Scan for a small set of repositories.
- Treat automations like workflow files. Decide who can create them, what they can change, and require human review before an automation's self-suggested changes are accepted.
- Pick review tiers per repository. Max for payment, authentication and infrastructure code; Lite or Balanced for the rest.
- On Enterprise Cloud, trial the package firewall with one organisation and a short cooldown on new npm versions, then measure how often builds get blocked.
ServiceNow hands the agent builder to employees
ServiceNow's EmployeeWorks update, announced 3 November 2026, has four parts. Productivity Assistants handle role-specific work with reasoning, memory and connections into enterprise tools, and employees configure them without code inside administrator-defined controls over behaviour, content, tools and access. Productivity Routines automate recurring work: describe an activity, choose the tools, deploy. ServiceNow's example is a sales agent that reviews upcoming meetings, researches prospects and prepares presentation material before the working day starts.
ToolSmith is the one developers should care about. It's a conversational development environment where you describe an automation in natural language and it plans and builds integrations, creates AI-agent tools, supports testing, and recommends fixes from logs. ServiceNow says it draws on knowledge of more than 75,000 APIs, and that it can also identify candidate workflows by reading operational logs.
That API index is the claim I'd want to verify. An index of 75,000 API surfaces tells you what exists and what it expects. It doesn't tell you that the endpoint your integration calls is deprecated, rate limited at 10 requests per second, or broken in the region you deploy to. Test ToolSmith output against a staging tenant before it touches anything that bills a customer.
Replit lets the support lead commit
Replit's Cloud Agents, available to all Pro and Enterprise customers, run a workflow that's more careful than the framing suggests. A subject matter expert describes a change in plain language. The agent locates the code, makes the change, and renders a live preview of the running application next to the chat. Before a pull request reaches a human, the agent opens the app in a browser and clicks through the change the way a customer would. If something fails, it fixes the code and retests. Screenshots and video of the final version go on the pull request. When a reviewer requests changes, the agent applies them and updates the request.
The browser click-through is the best idea in this batch of announcements. Visual verification as a precondition to review beats a screenshot pasted into Slack every time.
The self-fix loop needs a cap. An agent that retests until the browser stops complaining will eventually learn that a flaky selector is cheaper to work around than to fix. Give it a retry limit and make the retry count visible on the PR.
The resource nobody budgeted
Put these three together and you get a lot more pull requests. Copilot automations file work on a schedule. ServiceNow routines fire before the sales team wakes up. Replit agents open merge requests against your application. All of it lands in the same review queue, and reviewers were already the constraint.
GitHub's answer is tiering, which is sensible: Max for the repositories that touch money and auth, Lite for the CSS fixes. But code review Max doesn't help when the queue doubles, and it doesn't tell you which PRs were authored by an agent working from a routine someone configured six weeks ago and forgot.
Do these before the second repository gets wired up. Count agent-authored PRs separately in your metrics. Log automation creation as a change event with an owner attached. Cap retry loops in Replit and set an approval gate on self-modifying automations in GitHub. Then start with one repository, one automation, one review tier and a budget cap, and add the second repository once you know what the first one cost.