Agent Behavior
Rules for how the agent operates, tracks work, and communicates during coding sessions.
Activity Tracking
Every session must track what was done.
Log activity to docs/activity/ using your judgment on hierarchy:
docs/activity/ ├── 2025-01-15-feature-auth.md # By date + feature ├── 2025-01-15-bugfix-login.md # By date + type └── sessions/ └── 2025-01-15-session-1.md # By session
Activity Log Format
[Date] - [Brief Description]
What Was Done
- Bullet list of changes made
- Files modified
- Decisions made
Why
- Reasoning behind approach
- Trade-offs considered
What's Next
- Remaining work
- Known issues
- Questions for human
When to Log
Situation Action
Starting work Create/update activity log
Completing a task Summarize what was done
Making a decision Document the reasoning
Hitting a blocker Note the issue and questions
Ending session Final summary of state
Working Patterns
Ask vs. Proceed
Situation Action
Clear requirements Proceed
Multiple valid approaches Ask
Destructive operation Ask
Unclear scope Ask
Simple fix Proceed
Subagent Usage
Use subagents when:
-
Task is independent and parallelizable
-
Deep exploration needed without polluting main context
-
Multiple files need searching/analysis
Do directly when:
-
Simple, quick operation
-
Context is already loaded
-
Sequential dependency on previous work
Communication
-
Be concise - Don't over-explain obvious things
-
Show progress - Use todo lists for multi-step work
-
Surface blockers early - Don't spin on problems
-
Summarize at end - What was done, what's next
Quality Expectations
Before marking work "done":
-
Code runs without errors
-
Tests pass (if applicable)
-
Activity log updated
-
No obvious issues left unaddressed