Skip to content

Scheduled autonomous agents via cron + Claude Code #317

@williamjameshandley

Description

@williamjameshandley

Summary

Add scheduled autonomous agent capability by running Claude Code on a cron schedule with full MCP tool access. This bridges the gap between our current reactive system (user must message first) and the "scheduled autonomy" primitive emerging in products like Claude Cowork and Perplexity Computer.

Motivation

Our MCP architecture already provides:

  • Structured personal context (GTD cards, email, calendar)
  • Cross-referencing via links (urn:uuid:, mid:, calendar event IDs)
  • External LLM delegation
  • Mobile access via WhatsApp loop

The only missing piece is a triggering layer that runs agents on a schedule without user prompting.

Concrete use cases

  • Morning briefing — daily digest of calendar, due tasks, flagged emails
  • Auto-filing receipts — scan inbox for receipt emails, file to GTD
  • Payslip filing — Cambridge/Caius payslips arrive by email monthly, auto-file
  • Email triage — flag important unread, summarise overnight threads
  • Calendar prep — before meetings, pull relevant GTD cards and email threads
  • Recurring reviews — weekly GTD review, monthly finance check

Implementation sketch

A cron job running Claude Code with the existing MCP config:

0 8 * * * claude --prompt "Check GTD for tasks due today, process incoming receipts, summarise flagged emails"

Key design decisions

  1. Approval gates — some actions (sending emails, completing tasks) should queue for human approval rather than execute autonomously. The WhatsApp loop (mcp__loop) could serve as the approval channel.
  2. Error handling — unattended agents need graceful failure modes, logging, and cost limits.
  3. Scope control — each scheduled task should have a well-defined prompt and limited tool permissions.
  4. Cost management — token budgets per run to prevent runaway spending.

Context

Inspired by the "four new primitives of the agentic era" discussed in AI Daily Brief (26 Feb 2026): persistent work, multimodal orchestration, scheduled autonomy, and cross-device AI. Our system is strong on persistent state and cross-context access but weak on scheduled autonomy.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions