CHAPTERS
Workshop goal: turning Claude Code from a tool into a proactive teammate
Maya opens by framing the core problem: coding agents today are mostly reactive and wait for user prompts. The session’s goal is to show how “Routines” enable Claude Code to initiate work proactively, like a teammate who notices issues and acts.
Why running Claude Code on cron (or custom infra) is painful
The talk highlights why many existing approaches to proactive agents feel cumbersome. The overhead isn’t the prompt—it’s the infrastructure required to host, trigger, and supervise the agent reliably.
What Routines are: remote Claude Code automation with prompts + repos + triggers
Maya introduces Routines as a new Claude Code feature that packages the hard parts: remote execution, auth, state, and automation. You define what you want (instructions, repos, connectors, trigger), and Claude Code handles the rest.
Design pillars of Routines: always-on, proactive triggers, interactive sessions
Routines were built around three core product principles. These principles address the earlier pain points: reliability, flexible triggering, and real-time steerability.
Internal use case: automating documentation updates as PR volume increases
Maya explains why Anthropic needed automation for docs: Claude Code PR volume grew sharply, creating a bottleneck for the engineer maintaining documentation. Routines became a way to continuously keep docs in sync with product changes.
Creating a scheduled routine in the CLI with /schedule
The demo shows how a routine can be created from Claude Code via a natural-language schedule prompt. Claude asks follow-up questions (timing, notifications) and then creates the routine.
The three decisions for any routine: trigger, context, steerability
Maya outlines a reusable framework for designing routines. Each routine needs a clear trigger, sufficient context/tools, and a plan for how quality is ensured and the agent is guided.
Trigger options in depth: schedule-based vs event-based (GitHub + custom webhooks)
Routines support both time-based schedules and event-driven runs. Event triggers can be native GitHub events or custom events posted via an endpoint, enabling tight integration with developer workflows.
Context and connectors: repos, Drive/marketing materials, Slack notifications
Maya explains how to equip routines with the information and tools they need. Beyond code repos, routines can use connectors (e.g., Drive, Slack) to access reference content and notify stakeholders.
Steering and quality control: agent-on-agent review, live intervention, verification
The talk covers strategies to keep outputs reliable and aligned. Approaches include multi-agent review patterns, real-time human steering, and post-output verification like rendering docs.
Web UI walkthrough: inspecting a routine, sessions, and an auto-created docs PR
Maya demonstrates managing routines in claude.ai: viewing connected repos, triggers, and connectors. She opens a specific session to show how Claude compared repos, found changes, and created a PR.
Event-based demo: trigger on GitHub issue creation and steer the run
A second routine is set to fire when a GitHub issue is opened. The issue content is passed as context, Claude investigates whether it indicates a docs gap, and the session can be steered to stop if unnecessary.
Applying Routines to your workflows: deploy verifier and other proactive agents
Maya generalizes the pattern to common engineering scenarios. She outlines how to design a deploy-verifier routine with a deploy trigger, monitoring context, and a phased approach to autonomy.
Closing takeaways: proactive beats reactive; start with /schedule
Maya closes by reinforcing the value proposition: proactive agents reduce toil and react to problems without manual prompting. Routines are positioned as the infra-free way to operationalize these workflows quickly.
