Scheduled Agents

Agent steps work with schedule like any other step type. Add a schedule field to a DAG containing type: agent steps to run them on a cron schedule.

Prerequisites

• The scheduler must be running (dagu scheduler or dagu start-all)
• An agent model must be configured in Agent Settings (/settings/agent)
• See Scheduling for cron syntax and scheduler setup
• See Agent Step for agent configuration

Basic Example

A daily agent that summarizes system health:

schedule: "0 8 * * *"  # Every day at 8 AM

steps:
  - type: agent
    messages:
      - role: user
        content: |
          Check the system health by reviewing /var/log/syslog and
          report any errors or warnings from the last 24 hours.
    output: HEALTH_REPORT

Parameters

Parameters are defined at the DAG level. The same values apply to every scheduled run. Use ${VAR} syntax in message content:

schedule: "0 9 * * MON-FRI"

params:
  - LOG_DIR=/var/log/app
  - THRESHOLD=100

steps:
  - type: agent
    messages:
      - role: user
        content: |
          Scan all log files in ${LOG_DIR} for error counts.
          Flag any file with more than ${THRESHOLD} errors in the last 24 hours.
    output: ERROR_REPORT

To run with different parameter values, trigger the DAG manually or create separate DAG files.

Output Capture

The agent writes its result via the output tool, which is captured by the step's output field. Chain agent output to subsequent steps:

schedule: "0 6 * * *"

steps:
  - type: agent
    messages:
      - role: user
        content: "Analyze yesterday's deployment logs and summarize any issues."
    output: ANALYSIS

  - command: |
      echo "${ANALYSIS}" | mail -s "Daily Deployment Report" team@example.com

If the agent never calls the output tool, the output variable will be empty.

Catchup

When the scheduler restarts after downtime, catchup_window replays missed runs. Each missed interval triggers a full agent run, so be mindful of LLM API costs:

schedule: "0 * * * *"          # Hourly
catchup_window: "6h"           # Replay up to 6 hours of missed runs
overlap_policy: "skip"         # Skip if previous catchup run is still active

steps:
  - type: agent
    messages:
      - role: user
        content: "Check application metrics for the last hour and flag anomalies."
    output: MONITOR_RESULT

overlap_policy	Behavior
"skip" (default)	Drop the catchup run if the DAG is still running
"all"	Keep the run in the buffer, retry next tick
"latest"	Discard all but the most recent missed interval

Queue Control

Agent steps can be long-running and expensive (LLM API calls). Use queue to limit concurrent scheduled agent runs:

schedule: "*/30 * * * *"
queue: agent-tasks

steps:
  - type: agent
    messages:
      - role: user
        content: "Review open pull requests and summarize their status."
    output: PR_SUMMARY

Configure the queue in your Dagu config:

# ~/.config/dagu/config.yaml
queues:
  enabled: true
  config:
    - name: agent-tasks
      max_concurrency: 1  # Only one agent run at a time

Examples

Daily Log Analysis Agent

schedule: "0 7 * * *"

steps:
  - type: agent
    messages:
      - role: user
        content: |
          Analyze /var/log/app/errors.log from the last 24 hours.
          Categorize errors by type and severity.
          Suggest fixes for the top 3 most frequent errors.
    output: LOG_ANALYSIS

Agent Pipeline (Agent to Shell Step)

schedule: "0 9 * * MON"

steps:
  - type: agent
    messages:
      - role: user
        content: |
          Review the Git commit history for the past week.
          Write a concise weekly summary in Markdown format.
    output: WEEKLY_SUMMARY

  - command: echo "${WEEKLY_SUMMARY}" > /reports/weekly-$(date +%Y-%m-%d).md

Scheduled Agent with Restricted Tools and Bash Policy

schedule: "0 */6 * * *"

steps:
  - type: agent
    agent:
      tools:
        enabled:
          - bash
          - read
          - think
        bash_policy:
          default_behavior: deny
          deny_behavior: block
          rules:
            - name: allow-read-commands
              pattern: "^(cat|head|tail|grep|find|ls|wc|stat)\\b"
              action: allow
            - name: allow-network-checks
              pattern: "^(curl -s|ping -c|nslookup|dig)\\b"
              action: allow
      max_iterations: 30
    messages:
      - role: user
        content: |
          Run a security check:
          1. Check for world-readable files in /etc
          2. Verify no unexpected ports are open
          3. Check for failed SSH login attempts
    output: SECURITY_REPORT

Graph DAG with Multiple Agent Steps

schedule: "0 6 * * MON-FRI"
type: graph

steps:
  - name: check-infra
    type: agent
    messages:
      - role: user
        content: "Check disk usage, memory, and CPU across all servers."
    output: INFRA_STATUS

  - name: check-app
    type: agent
    messages:
      - role: user
        content: "Review application health endpoints and response times."
    output: APP_STATUS

  - type: agent
    messages:
      - role: user
        content: |
          Combine these reports into a single morning briefing:

          Infrastructure: ${INFRA_STATUS}
          Application: ${APP_STATUS}
    depends: [check-infra, check-app]
    output: MORNING_BRIEF

The check-infra and check-app steps run in parallel. The summarize step waits for both to finish before combining their outputs.

See Also

• Scheduling — Cron syntax, catchup, overlap policy, queue management
• Agent Step — Agent configuration, tools, bash policy, output capture
• Tools Reference — Full parameter documentation for each tool
• Memory — Persistent agent memory across sessions