Deep Work Session

Deep Work Session

A Deep Work session is a focused period where an AI agent works on a specific Agent Hive project. This skill guides you through the complete session lifecycle.

Session Lifecycle

1. ENTER -> 2. CLAIM -> 3. WORK -> 4. UPDATE -> 5. HANDOFF

Starting a Deep Work Session

Via Dashboard

1. Run the dashboard:

   make dashboard
   # Or: uv run streamlit run src/dashboard.py
   

2. Open http://localhost:8501

3. Select a project from the sidebar

4. Click "Generate Context"

5. Copy the context and paste to your AI agent

Via Script

make session PROJECT=projects/your-project
# Or: ./scripts/start_session.sh projects/your-project

Manual Entry

Read the AGENCY.md file directly and follow the protocol below.

Your Responsibilities

When entering a Deep Work session, you MUST:

1. Read AGENCY.md first - Understand the project context fully
2. Claim the project - Set owner to your agent name
3. Work on priority tasks - Focus on highest priority incomplete items
4. Update progress - Mark tasks complete, add notes
5. Handle blocking - Set blocked: true if you need help
6. Follow handoff protocol - Clean up state before ending

Claiming a Project

Before starting work, verify and update the frontmatter:

# Check these are true:
status: active        # Must be active
blocked: false        # Must not be blocked
owner: null           # Must be unclaimed

# Then set owner:
owner: "claude-sonnet-4"
last_updated: "2025-01-15T14:30:00Z"

Working on Tasks

Task Prioritization

Work on tasks in this order:

1. Tasks marked critical priority
2. Tasks marked high priority
3. Tasks that unblock other projects
4. Remaining tasks by document order

Marking Progress

Update the markdown content as you complete tasks:

## Tasks
- [x] Research existing solutions    # Completed
- [x] Design architecture            # Completed
- [ ] Implement core feature         # In progress
- [ ] Write tests                    # Not started

Adding Agent Notes

Document your work with timestamped notes:

## Agent Notes
- **2025-01-15 15:45 - claude-sonnet-4**: Completed architecture design.
  Chose event-driven pattern for scalability. Implementation ready to begin.
- **2025-01-15 14:30 - claude-sonnet-4**: Starting research phase.
  Will evaluate 3 existing solutions.

Handling Blockers

If you encounter something you cannot resolve:

blocked: true
blocking_reason: "Need database credentials from DevOps team"
status: blocked  # Optional: change status too

Add a note explaining:

- **2025-01-15 16:00 - claude-sonnet-4**: BLOCKED - Cannot proceed with
  database integration. Need credentials from DevOps. Created ticket #123.

Handoff Protocol

Before ending your session, you MUST complete these steps:

1. Update All Completed Tasks

Mark everything you finished with [x]

2. Update Timestamp

last_updated: "2025-01-15T16:30:00Z"

3. Add Final Notes

Document what was accomplished and any context for the next agent:

- **2025-01-15 16:30 - claude-sonnet-4**: Session complete. Finished
  research and design phases. Implementation is ready to begin - start
  with `src/feature.py`. Note: the authentication module has a known
  issue (see issue #45).

4. Release or Retain Ownership

If completely done with the project:

owner: null
status: completed  # If all tasks done

If handing off to another agent:

owner: null
status: active  # Keep active for next agent

If you'll continue later:

owner: "claude-sonnet-4"  # Keep ownership
# Add note about when you'll return

5. Set Blocking if Needed

blocked: true  # Only if external help needed
blocking_reason: "Waiting for API access approval"

Generated Context Format

The Dashboard generates this context package:

# DEEP WORK SESSION CONTEXT
# Project: project-id
# Generated: 2025-01-15T14:30:00

---

## YOUR ROLE
[Responsibilities and instructions]

---

## AGENCY.MD CONTENT
[Full AGENCY.md file with frontmatter and content]

---

## PROJECT FILE STRUCTURE
[Directory tree of project files]

---

## AVAILABLE COMMANDS
[What actions you can take]

---

## HANDOFF PROTOCOL
[Required steps before ending]

---

## BOOTSTRAP COMPLETE
[Signal to begin work]

Best Practices

1. Claim immediately - Prevent conflicts with other agents
2. Update frequently - Keep state current throughout session
3. Be thorough in notes - Future agents need context
4. Don't overcommit - Better to complete fewer tasks well
5. Clean handoff - Always follow the handoff protocol
6. Use coordinator - If available, use real-time coordination server

Session Checklist

Starting

• Read AGENCY.md completely
• Verify project is claimable
• Set owner to your agent name
• Update last_updated
• Identify highest priority task

During

• Mark tasks as you complete them
• Add notes for significant progress
• Set blocked if you hit a wall
• Update last_updated periodically

Ending

• All completed tasks marked with [x]
• Final last_updated timestamp
• Comprehensive closing notes added
• owner released or retained intentionally
• status updated if appropriate
• blocked/blocking_reason set if needed