Team collaboration with AI
Once a whole team works through agents, consistency becomes the problem. Two engineers asking two agents the same question shouldn't get two architectures. The fix is shared configuration, clear review norms, and decisions recorded where agents read them.
Multi-developer AI workflows
Shared configuration:
project/
├── CLAUDE.md # Team conventions (committed)
├── .mcp.json # Team MCP servers (committed)
├── .claude/
│ ├── settings.json # Shared hooks and permissions (committed)
│ ├── commands/ # Team slash commands (committed)
│ ├── skills/ # Team skill library (committed)
│ ├── agents/ # Team subagents (committed)
│ └── memory/ # Architecture and decisions (committed)
└── CLAUDE.local.md # Personal preferences (gitignored)Team CLAUDE.md practices:
- Record architectural decisions so every agent gives consistent advice
- NEVER rules for dangerous operations
- Approved libraries and patterns, with reference files to copy
- Owned via CODEOWNERS, changed via PR
Team norms to write down:
| Norm | Example |
|---|---|
| Who approves AI PRs | The domain owner via CODEOWNERS, never only the author |
| PR size | Agents make huge diffs easy. Cap them. Split by behavior |
| Labeling | Mark PRs that are mostly agent-generated, so reviewers know what to look for |
| Isolation | Worktrees per task, feature branches, never direct to main |
| Lock files | Committed and reviewed like code |
Code review for AI-generated PRs
Team review checklist:
## AI Code Review Checklist
### Correctness
- [ ] Logic matches the ticket
- [ ] Edge cases handled
- [ ] No obvious bugs
### Security
- [ ] No hardcoded secrets
- [ ] Input validation at trust boundaries
- [ ] OWASP top 10 considered
### Fit
- [ ] Follows existing patterns
- [ ] No unnecessary abstractions
- [ ] Naming matches team standards
### Tests
- [ ] Tests cover the change
- [ ] Tests assert behavior, not just coverage
- [ ] No flaky tests introduced
### AI-Specific
- [ ] No hallucinated imports or dependencies
- [ ] No invented APIs
- [ ] Comments are accurateReview prompt:
"Review this PR from another engineer's agent session. Check:
- Does it follow src/services/ patterns?
- Any security issue we should block on?
- Is the complexity justified?"Watch review load. Agents increase PR volume. If it lands on the same two reviewers, they burn out and review quality drops. Spread it deliberately and measure it.
Onboarding new team members
New engineers can learn the codebase from the same context the agents use:
"Read CLAUDE.md and .claude/memory/ARCHITECTURE.md. Summarize:
1. The tech stack
2. Key architectural patterns
3. Things I should never do
4. How to run the dev environment"For an unfamiliar codebase, a generated knowledge graph such as Understand-Anything helps too.
Juniors still need to build judgment. If they only ever prompt, they never learn to evaluate. Rotate them through code review, debugging, and incident response, and ask them to explain agent-written diffs in their own words.
Handling AI disagreements
When agents give different team members different answers:
- Decide once, in writing. An ADR, not re-prompting until someone's answer wins
- Record it where agents read it: CLAUDE.md or
.claude/memory/ - Encode workflows as skills so everyone runs the same process
- Cross-check important decisions with a fresh session or a second model