Night Shift — Autonomous Overnight Work

Continuous autonomous work mode. The agent works through the night, spawning parallel sub-agents and processing results as they complete.

When to Use

User says: "night shift", "work through the night", "autonomous mode", "keep working while I sleep"

Architecture

code

Heartbeat-driven continuous loop:

Main Session (Orchestrator/Opus):
  → Reads state.json, spawns 2-4 sub-agents IN PARALLEL
  → As sub-agents complete (announce messages), immediately:
     1. Update state.json
     2. Spawn replacement sub-agents for next tasks
     3. Repeat until queue empty
  → Cron is ONLY a safety net — restarts work if agent dies

Sub-Agents (Workers/Sonnet):
  → Focused single task, fresh context
  → Write output to project dir
  → Report back automatically via announce

Why Parallel

Sequential sub-agents = 1 task per cycle = slow. Always spawn 2-4 sub-agents at once.

•Sub-agents are independent — no coordination needed between them
•Each takes 2-5 minutes; running 3 in parallel = 3x throughput
•sessions_spawn supports up to 4 concurrent, 8 max total
•Stagger slightly if they share resources (web_search rate limit)

Safety Net Cron

One cron job ensures the agent keeps working even if it gets stuck or the session dies:

code

schedule: { kind: "every", everyMs: 1200000 }  # 20 min
payload: { kind: "systemEvent", text: "Night shift safety check: Read projects/night-shift-YYYY-MM-DD/state.json. If there are pending tasks and fewer than 2 sub-agents running, spawn more. If all done, compile summary." }
sessionTarget: "main"

This is NOT the main work driver — it's a fallback. The main loop is driven by sub-agent completions triggering the next wave.

Execution Flow

code

1. Lucas says "activate night shift"
2. Create state.json with full task queue
3. Set safety cron (20 min)
4. Spawn 2-4 sub-agents immediately (first wave)
5. Sub-agent A completes → announce → spawn replacement
6. Sub-agent B completes → announce → spawn replacement
7. ... continuous pipeline until queue empty
8. Compile summary
9. Lucas wakes → remove cron, deliver report

Model Strategy

•Opus: Orchestrator (main session), architecture/design, implementation decisions
•Sonnet: Research tasks, web scraping, file processing, routine work
•Set per-spawn with model parameter in sessions_spawn

Sub-Agent Prompt Guidelines

Keep prompts focused and minimal:

•One clear objective per spawn
•List specific tools and search queries
•Specify exact output file path
•Do NOT include full workspace context — they inherit tools
•Add 2s delays between web_search calls (Brave rate limit: 1 req/sec)
•Use label parameter for tracking (e.g., label: "vtuber-research-1")

Delegation to Other Grid Agents

For tasks outside your role (work stuff → Robson), use bot-bridge:

bash

~/openclaw/scripts/bot-bridge.sh robson "Night shift task: [description]. Write output to ~/openclaw/projects/night-shift-YYYY-MM-DD/[filename].md"

Send all Robson tasks at the START of the night, not sequentially. He works in parallel too.

Activation Sequence

1. Create Output Directory

bash

mkdir -p projects/night-shift-YYYY-MM-DD/

2. Initialize State File

Write projects/night-shift-YYYY-MM-DD/state.json:

json

{
  "date": "2026-01-31",
  "startedAt": "2026-01-31T22:00:00-03:00",
  "status": "active",
  "taskQueue": [
    {"id": "t01", "name": "Task Name", "status": "pending", "model": "sonnet"},
    {"id": "t02", "name": "Task Name", "status": "pending", "model": "sonnet"}
  ],
  "currentTask": null,
  "runningTasks": [],
  "completedTasks": [],
  "findings": []
}

3. Set Safety Cron

code

schedule: { kind: "every", everyMs: 1200000 }
sessionTarget: "main"
payload: { kind: "systemEvent", text: "Night shift safety check..." }

4. Spawn First Wave (2-4 sub-agents)

Immediately spawn multiple sub-agents. Don't wait for the first cron tick.

code

sessions_spawn(task="...", model="sonnet", label="task-name-1")
sessions_spawn(task="...", model="sonnet", label="task-name-2")
sessions_spawn(task="...", model="sonnet", label="task-name-3")

5. On Each Sub-Agent Completion

When an announce message arrives:

•Update state.json (mark completed, add findings)
•If pending tasks remain → spawn next sub-agent immediately
•If all done → compile summary, notify if urgent
•Track runningTasks array to avoid over-spawning (max 3-4 concurrent)

6. On Safety Cron Tick

•Check state.json
•If runningTasks < 2 and pending tasks exist → spawn more
•If all tasks done → compile summary and remove cron
•If nothing to do → HEARTBEAT_OK

Task Types

Research Tasks

Model: Sonnet (cheap, sufficient). Tools: web_search → web_fetch → Jina Reader (r.jina.ai/URL) → browser (last resort). Output: markdown file with TL;DR, Key Findings, Action Items, Sources.

Build Tasks

Model: Opus for architecture, Sonnet for implementation. Rules: branch (night-shift/YYYY-MM-DD-topic), never push to main.

Processing Tasks

Obsidian vault, ReadItLater items, email digests. Output: digest file + action items.

Analysis Tasks

Review previous outputs, synthesize patterns, identify opportunities. Run after research rounds to decide next steps.

Output Structure

code

projects/night-shift-YYYY-MM-DD/
├── 00-plan.md                    # Task list and goals
├── 01-[task-name].md             # Individual task outputs
├── 02-[task-name].md
├── ...
├── summary.md                    # Morning summary
└── state.json                    # Night shift state

Default Research Topics

Customize per night, but defaults:

•Efficiency — Token optimization, faster workflows, better tools
•New Skills & Tools — GitHub trending, new AI tools, CLI utilities
•Multi-Agent / Swarms — Agent orchestration, communication patterns
•Virtual Presence — Avatars, Meet bots, screen recording pipelines
•AI Agent Trading & Crypto — Trading bots, DeFi, Polymarket
•Security Assessment — SSH, Tailscale, firewall, secrets
•Obsidian Vault / ReadItLater — Process saved items, organize knowledge

Rules

Must Follow

•Drop everything if Lucas messages — respond immediately
•Parallel by default — always 2-4 sub-agents running
•One file per task — never dump everything into one file
•Track running tasks — update state.json on every spawn/completion
•Follow workspace structure — output in projects/, not memory/

Should Follow

•Rotate categories — mix research, build, and processing tasks
•Front-load Robson delegation — send ALL his tasks at night start
•If a task fails twice, mark it and move on
•Prefer web_search + web_fetch over browser (faster, cheaper)
•Use browser for dynamic sites (X/Twitter, authenticated pages)
•Label all sub-agents for tracking (sessions_spawn label param)

Morning Summary

When user wakes up, compile summary.md:

markdown

# 🌅 Night Shift Report — [Date]

## Highlights
- [Top 3-5 discoveries, one line each]

## Tasks Completed
| # | Task | Status | Output |
|---|------|--------|--------|

## Key Findings
- [Finding]: [Why it matters]

## Action Items
- [ ] [Next steps based on findings]

## Built Tonight
- [What was implemented, branch name]

## Process Notes
- Tasks completed: X/Y
- Parallelization: avg N concurrent sub-agents
- Improvements for next time: [lessons]

Deactivation

Night shift ends when user says to stop. On deactivation:

•Compile summary.md
•Remove safety cron
•Update state.json status to "completed"
•Send morning summary
•Reset HEARTBEAT.md