Vex is a coding orchestrator built on Vellum. It has its own GitHub bot account, its own Slack identity, and a recurring schedule that keeps running whether the human guardian is awake or not. Instead of waiting for someone to ask, Vex watches the engineering surface area and brings problems forward while they are still fresh.
What Vex watches
Vex scans Slack, GitHub, CI, and Sentry on a cadence. It looks for the signals that usually get scattered across tools: failing checks, auth regressions, broken desktop clients, crash spikes, suspicious PRs, and unresolved incident threads. When something looks wrong, it gathers the context and turns the alert into an engineering workflow.
Run engineering watch loop every 15 minutes.
Check Slack, GitHub, CI, Sentry, and Linear for signals that something is broken or stuck. Look for production regressions, failed checks, crash spikes, auth issues, unresolved incident threads, risky PRs, and bugs that need follow-up.
When you find something important:
1. Gather the relevant Slack, GitHub, CI, Sentry, and Linear context
2. Summarize the likely impact and blast radius
3. Create or update the Linear ticket if needed
4. Review related PRs and request changes when appropriate
5. Dispatch Claude Code, Codex, or Devin if implementation work is needed
6. Write a short incident or postmortem-style summary
7. Post the team-facing update with links and next steps
Keep memory of what you saw in prior runs so you can track what changed, what is still unresolved, and what needs escalation.How it works
The schedule skill lets Vex run repeatedly, for example every 15 minutes, without waiting for a manual prompt. Each run can reuse context from prior runs, so Vex is not starting from zero. It remembers what it saw before, what changed, and which issues are still unresolved.
When Vex finds a code problem, it can review the PR, request changes, or dispatch another coding agent. Through coding agent orchestration, it can hand a focused task to Claude Code or Codex. Through external engineering tools, it can coordinate with systems like Devin when a bug needs deeper implementation work.
When the problem is operational, Vex can turn Sentry crashes into Linear tickets, connect the issue to relevant Slack and GitHub context, and write a postmortem-style summary. The value is not just that Vex sees the alert. It understands enough of the surrounding codebase and team workflow to make the alert actionable.
Why this pattern matters
Most engineering automation waits for a trigger: a person files a ticket, a CI job fails, someone asks a bot for help. Vex is different because it behaves like a persistent teammate. It has an identity in the tools, a memory of the codebase, and a loop that keeps checking the system even when no one is actively watching.
That changes the shape of engineering work. Bugs do not have to sit in Slack until morning. PRs do not have to wait for a human to notice a failed check. Incidents can arrive with context already attached. The assistant becomes the connective tissue between monitoring, code review, issue tracking, and coding agents.
The outcome
The team wakes up to a prepared engineering brief instead of a pile of disconnected alerts. Vex has already scanned the tools, documented the issue, routed the work, and preserved the context. The loop is real, and it keeps running.