Skip to content

research-git-native-agents.md

Latest commit

 

History

History
208 lines (167 loc) · 8.29 KB

research-git-native-agents.md

File metadata and controls

208 lines (167 loc) · 8.29 KB

Git-Native Agent Cooperation: The Complete Playbook

"It's not local or sandbox, it's git-native." — Captain Casey

Executive Summary

Current AI coding tools (Aider, Claude Code, Cursor, Crush) are single-agent, local-first, chat-driven. Our fleet operates as multi-agent, repo-first, git-native. This document captures the research, architecture, and specific GitHub exploitation strategies to make our fleet the most effective agentic development system ever built.

The Paradigm Shift

Dimension Old (Aider/Claude/Cursor) New (FLUX Fleet)
Agents 1 N (fleet)
Coordination Chat (synchronous) Git commits (asynchronous)
State Local files GitHub repos
CI/CD Manual Free (GitHub Actions)
Verification Self-test Cross-agent adversarial review
Knowledge Session context Repo history (permanent)
Scaling Bigger model More agents
Human role Prompt engineer Captain (approve/redirect)
Cost $200/month per tool Free (GitHub free tier)

5 Killer Advantages Over Single-Agent Tools

1. Adversarial Verification (Red Team / Blue Team)

  • WriterAgent writes code
  • TestAgent writes adversarial tests designed to break it
  • ReviewAgent checks architectural compliance
  • Result: 3x more edge cases caught than self-testing

2. Parallelization by Architectural Boundary

  • RouterAgent parses task into dependency graph
  • Multiple agents work different branches simultaneously
  • Merge when tests pass independently

3. Permanent Institutional Memory

  • Every decision is a commit with metadata
  • Agent diaries in repos, not session contexts
  • New agents bootstrap from repo history

4. Free CI/CD at Scale

  • GitHub Actions: 2,000 min/month free
  • Auto-test on every push
  • Artifacts stored for agent review
  • Cross-repo notification via dispatch events

5. Captain-Not-Bottleneck

  • Human approves merges, doesn't write code
  • Agent trust scores auto-promote reliable agents
  • Escalation paths for blocked tasks
  • Dashboard (oracle1-index) shows fleet status at a glance

30 GitHub Features Exploited by Agents

Compute (Free!)

# Feature Agent Usage A2A Value
1 Actions Run tests, builds, analysis on push Automatic verification
2 Codespaces Spin up dev environments Isolated workspaces
3 Packages Publish/share libraries Component sharing
4 Artifacts Store test results, build outputs Agent-to-agent handoff
5 Pages Host dashboards, docs Fleet visibility

Communication

# Feature Agent Usage A2A Value
6 Issues Create/assign/track tasks Asynchronous work queue
7 PRs Propose changes, request review Structured collaboration
8 Discussions Debate, plan, share knowledge Consensus building
9 Webhooks Real-time event notification Event-driven coordination
10 Notifications @mention, subscription alerts Attention routing

Organization

# Feature Agent Usage A2A Value
11 Projects v2 Kanban boards, status tracking Fleet kanban
12 Labels State machine (status, priority, domain) Task routing
13 Branches Isolated work streams Parallel execution
14 Forks Copy-modify-contribute pattern Safe experimentation
15 Tags Version milestones Coordinated releases

Quality & Security

# Feature Agent Usage A2A Value
16 Checks (Status API) Pass/fail gates on commits Multi-agent approval
17 Code Scanning Automated security analysis Coordinated vuln management
18 Dependabot Auto-dependency updates Supply chain coordination
19 Branch Protection Mandatory reviews Governance automation
20 Merge Queues Ordered PR merging Conflict prevention

Knowledge

# Feature Agent Usage A2A Value
21 Wikis Collaborative documentation Protocol specs
22 Releases Version communication Downstream coordination
23 Gists Share snippets/configs Quick info exchange
24 GraphQL API Efficient complex queries Informed decisions
25 Insights Analyze activity patterns Workload distribution

Infrastructure

# Feature Agent Usage A2A Value
26 Environments Staged deployments Coordinated rollouts
27 Secrets Secure credential storage Authenticated ops
28 Deployments Track what's deployed where Audit trail
29 Templates Quick repo spawning Standardized setup
30 Orgs/Teams Permission structure Role-based access

Fleet CI/CD Architecture

Workflow 1: Auto-Test (Triggered by Agent Commits)

Agent pushes → GitHub Actions → detect language → run tests → artifact upload
If PR: → create check status → reviewer agent sees green/red

Workflow 2: Cross-Repo Sync (Agent-to-Agent Notification)

Agent A pushes to main → dispatch event → Agent B's repo gets notified
→ Agent B checks if it needs to update

Workflow 3: Fleet Health (Scheduled)

Every 30 min → scan fleet repos → check CI status → update dashboard
→ flag stale repos → create issues for attention

Workflow 4: TODO-to-Issue (Knowledge Mining)

Every 6h → scan TODO/FIXME → create issues → auto-assign by label
→ agents claim from issue queue

Free Tier Budget

Resource Free Allowance Fleet Usage
Actions minutes 2,000/month ~200 repos × 5 min = 1,000
API requests 5,000/hour Agents poll every 5 min = 600/hr
Storage 500MB Test artifacts, reports
Pages Unlimited Dashboard hosting
Codespaces 120 core-hours Agent runtime
Packages 500MB Library distribution

Total cost: $0/month vs $200+/month for single-agent tools.

The Kimi Architecture: Git-Native Agent Swarm (GNAS)

┌─────────────────────────────────────────────────────────────┐
│                    FLEET ORCHESTRATOR                        │
│         (GitHub App: "FleetCommand" running on Actions)      │
└──────────────┬──────────────────────────────────────────────┘
               │ Webhooks / GraphQL
    ┌──────────┴──────────┐
    ▼                     ▼
┌─────────┐         ┌──────────┐
│ AGENT   │◄───────►│  AGENT   │   ... (N agents)
│   REPO  │  PR/Iss │   REPO   │
│ (Micro- │         │ (Micro-  │
│  service)│         │  service)│
└────┬────┘         └────┬─────┘
     │                   │
     └───────┬───────────┘
             ▼
    ┌─────────────────┐
    │  SHARED CANON   │
    │  (Target Repo)  │  ← The actual product codebase
    └─────────────────┘

Agent Structure: Each agent is a repo containing:

  • src/: Agent's own logic (self-modifying)
  • prompts/: System prompts versioned via git
  • tools/: Specific capabilities
  • .github/workflows/: Agent's runtime

Coordination Protocol:

  • Issues = Task contracts (structured YAML frontmatter)
  • PRs = Proposals with semantic diffs
  • Commits = Checkpointing reasoning
  • Labels = State machine

Research Sources

  • Seed (SiliconFlow): Creative ideation on GitHub exploitation
  • Kimi-K2.5 (SiliconFlow): Big-picture architecture, 5 killer advantages
  • DeepSeek-V3.2 (SiliconFlow): 30-feature GitHub catalog
  • Oracle1 🔮: Synthesis and implementation

Part of the FLUX Fleet Research Series. Generated 2026-04-11.