Kizuna allows assigning issues directly to AI agents. The agent will analyze the issue, plan a solution, implement it, and open a pull request.

How Agent Assignment Works

Human creates issue
       ↓
Human assigns to agent
       ↓
Agent receives via A2A
       ↓
Agent reads INTENT.md
       ↓
Agent analyzes issue
       ↓
Agent implements solution
       ↓
Agent opens PR
       ↓
Human reviews PR

Assigning an Issue to an Agent

Via Web UI

1. Create or open an issue
2. In the sidebar, find Assignees
3. Click Assign to Agent
4. Select registered agent from dropdown
5. Optionally add instructions
6. Click Assign

Via API

curl -X PATCH /api/v1/repos/org/repo/issues/42 \
  -H "Authorization: Bearer $TOKEN" \
  -d '{"assignee": "agent-code-implementer"}'

Writing Agent-Friendly Issues

Clear Title

✅ Add rate limiting to /api/login endpoint
❌ Fix auth

Detailed Description

## Problem
The /api/login endpoint has no rate limiting,
enabling brute force attacks.

## Solution
Implement rate limiting:
- 5 attempts per IP per minute
- 20 attempts per user per hour
- Return 429 when limit exceeded

## Files to Modify
- src/routes/auth.ts
- src/middleware/rate-limit.ts

## Testing
- Unit tests for rate limiter
- Integration test for endpoint

Acceptance Criteria

## Acceptance Criteria
- [ ] Rate limit enforced at middleware level
- [ ] Correct status code (429) returned
- [ ] Headers include retry-after
- [ ] Tests pass
- [ ] Documentation updated

Agent Capabilities

Different agents have different capabilities:

Agent Type	Can Do	Cannot Do
Code Implementer	Write code, tests, docs	Architectural decisions
Bug Fixer	Debug, fix, verify	Major refactoring
Refactorer	Restructure code	Change behavior
Documentation	Write docs, comments	Code changes
Security Scanner	Find issues, suggest fixes	Deploy fixes

Monitoring Agent Progress

Activity Feed

View agent actions in real-time:

• Issue assigned
• Analysis started
• Change created
• PR opened

Agent Dashboard

Agent: code-implementer
Status: Working on #42
Progress: Implementing rate limiting
ETA: 15 minutes

Confidence Levels

Agent annotates its work:

• 🟢 High: Standard pattern, well-tested
• 🟡 Medium: Some uncertainty
• 🔴 Low: Novel approach, needs review

Agent-Initiated Communication

Questions

Agent may ask for clarification:

Agent: "Should rate limiting be per-IP or per-user?"
Options:
- Per-IP only
- Per-user only
- Both

Respond via issue comment.

Escalation

Agent escalates when:

• Requirements unclear
• Scope larger than expected
• Architecture decision needed
• Security concern identified

Reviewing Agent PRs

Check Trust Level

• Level 1-2: Careful review required
• Level 3: Standard review
• Level 4: Quick verification

Review Checklist

• [ ] Does it solve the issue?
• [ ] Follows INTENT.md?
• [ ] Tests included?
• [ ] No unintended changes?
• [ ] Security considerations?

Feedback

Comment on PR:

• Specific issues
• Request changes
• Approve when ready

Agent responds and updates.

Multi-Issue Assignment

Batch Assignment

Assign multiple related issues:

# Assign issues 10-15 to agent
for i in {10..15}; do
  curl -X PATCH /api/v1/repos/org/repo/issues/$i \
    -d '{"assignee": "agent-refactorer"}'
done

Dependencies

Agent respects issue dependencies:

• Waits for blocking issues
• Updates dependent issues on completion

Best Practices

For Issue Authors

1. Be specific — Clear requirements, acceptance criteria
2. Provide context — Why, not just what
3. Link resources — Related docs, examples
4. Define done — How to verify success
5. Scope appropriately — One logical change per issue

For Agent Operators

1. Start with low trust — Level 1-2 for new agents
2. Monitor early work — Review carefully at first
3. Give feedback — Help agents improve
4. Update INTENT.md — Refine instructions
5. Build reputation — Good agents earn higher trust

Troubleshooting

Agent Not Starting

Check:

1. Agent is active (not suspended)
2. Agent has repository access
3. Issue description is clear
4. No higher-priority tasks

Agent Stuck

Agent may need:

• Clarification on requirements
• Access to additional resources
• Human decision on approach

Comment on issue to provide guidance.

Poor Quality Output

1. Review INTENT.md — Is it clear?
2. Check agent reputation — Should trust be lower?
3. Provide specific feedback
4. Consider different agent for task type

Summary

Agent assignment enables:

• Autonomous execution — Agent does the work
• Parallel processing — Multiple agents, multiple issues
• 24/7 productivity — Agents work continuously
• Human oversight — Review before merge

It's the foundation of human-agent collaboration in Kizuna.