agent-swarming

Multi-Agent Collaboration & Swarming

Cross-cutting pattern for Orryx Group development. Originally Triora's; promoted to cross-cutting in Wave 3 of the architecture migration.

Supports multi-agent swarming between Claude and ChatGPT for autonomous development.

Collaboration Principles

• Assume other agents will consume outputs — use clear headings, lists, JSON/YAML
• Behave collaboratively and critically — acknowledge useful work, flag inconsistencies
• Cross-check critical logic — recommend independent re-derivation for high-stakes outputs (matching logic, financial calculations, security policies, etc.)
• Swarming for autonomy — agents work together with minimal human intervention

Agent Roles in Development Swarm

Agent	Model	Primary Role	Tier source
Engineer	Claude	Code implementation, testing	sonnet
Architect	Claude	Design, architecture decisions	opus
Tester	Claude	Test generation, coverage	sonnet
Researcher	Claude Opus	Deep research, knowledge gathering	opus
Designer	Claude	UI/UX design, accessibility	sonnet
Hygiene	Claude	Repo cleanliness, archival, consolidation	sonnet
Reviewer	GPT-4	Code review, verification	n/a (external)
QA	GPT-4	Integration testing, acceptance	n/a (external)
Explorer	Claude	Read-only investigation	haiku, permissionMode: plan
Doc-writer	Claude	Documentation generation	haiku

Tier assignments are canonical via foundation/model-tiering-convention. Configs typically live in agents/claude/<role>.yaml and agents/chatgpt/<role>.yaml per repo.

Parallel Research Agent

The Researcher Agent runs independently in parallel to:

• Conduct deep research on regulatory, technical, and domain topics
• Produce cited markdown reports in research/outputs/
• Inform architecture and compliance decisions
• Use Graph-of-Thoughts methodology for complex topics

Launch: select a prompt, refine, start a new Claude Code session with /model opus, execute research while main agent continues.

Cross-Model Verification

For critical outputs (especially regex / async / security / matching logic):

Claude (generate) → GPT-4 (verify) → Compare
  Agreement → High confidence, proceed
  Minor disagreement → Iterate and refine
  Major disagreement → Escalate to human

See foundation/cross-model-verification for the formal protocol.

Agent Handoff Protocol

## Agent Handoff: [Task ID]
### Context
- Task: [Description]
- Phase: [Current phase]
- From: [Agent ID]  /  To: [Agent ID]
### Completed Work
- [x] What was done  /  Files changed
### Verification Requested
- [ ] Correctness  /  Security  /  Tests
### Confidence: [0.0-1.0]

Swarming Patterns

Pattern	Description	Use Case
Sequential	A → B → C	Standard feature development
Parallel	[A + B] → Merge	Independent components
Verification	A → B (verify)	Critical decisions
Debate	A + B → Arbitrate	Architecture choices
Refinement	A → B (review) → A (fix) → Loop	Complex logic

⚠️ Branch Isolation Rule (ENFORCED)

Parallel agents MUST work on separate branches. This is non-negotiable.

Why

Concurrent agents on the same branch cause commit collisions. This was learned the hard way in Triora when a compliance agent stepped on the engineer's commits, resulting in lost work and a messy force-push recovery.

Rules

1. One agent per branch — before starting, check git branch and ensure no other agent targets your branch
2. Branch naming must identify the agent — use patterns like feat/agent-name-task or fix/agent-name-issue
3. Never force-push a shared branch — if conflicts arise, rebase locally and resolve
4. Coordinate merges — agents merge to main sequentially, not simultaneously
5. Check before commit — run git status and git log --oneline -3 before committing

Pre-Commit Check

current_branch=$(git branch --show-current)
echo "Committing to: $current_branch"
git status

If Collision Detected

1. STOP immediately — do not force-push
2. Stash: git stash
3. Pull: git pull --rebase
4. Re-apply: git stash pop
5. Resolve conflicts manually
6. Report the collision to the coordinating agent/human

See also

• orchestration/cpd-protocol — Continuous Parallel Development (5-stream pattern)
• orchestration/ralph-loop — autonomous-loop pattern
• orchestration/git-worktrees — workspace isolation pattern
• foundation/model-tiering-convention — which tier for which role
• foundation/cross-model-verification — the verifier handoff schema
• Triora original: D:\Clinical.Trials\.claude\skills\agent-swarming\SKILL.md