Multi-agent integration patterns
What this lesson covers
Section titled “What this lesson covers”L14 takes the worktree foundation from L13 and covers the patterns for integrating work from parallel AI agents. Three patterns: shared origin (the default for most teams), per-agent fork (strongest isolation, for untrusted or external agents), and shared worktrees (fastest iteration for solo leads).
The lesson covers the lead orchestrator’s role in detail: planning, spawning, watching, integrating, catching semantic conflicts, push gate, cleanup. It also introduces the most important failure class in multi-agent work: SEMANTIC conflicts that git cannot detect because they cross file boundaries without producing line-level conflicts.
The lesson references the Clawless 2026-06-04 multi-agent sprint as a real-world example of how a lead-stage build caught ~300 latent breakers no per-agent test had found.
L14 sets up L15 (AI-authored commits and PRs) and L16 (the future of git in an AI world).
By the end of L14, the reader will be able to
Section titled “By the end of L14, the reader will be able to”- Distinguish between shared-origin, per-agent fork, and shared-worktrees integration patterns
- Choose the right pattern for a given situation using the decision tree (need other-human review, untrusted agents, fast solo iteration, default)
- Play the lead orchestrator role: plan, spawn, watch, integrate, catch semantic conflicts, push gate, cleanup
- Recognize semantic conflicts as a distinct class from git-level conflicts and apply guardrails (tight agent scope, integration tests, lead-stage build, diff review)
- Apply the lead-stage build as the most important guardrail in multi-agent integration
- Decide when to abort and re-run an agent vs repair its branch by hand
Prerequisites
Section titled “Prerequisites”L1-L13 of T7. Especially: branches (L5), merges (L7), cherry-pick (L11), rebase (L12), and worktrees (L13). L14 builds on all of these in tight combination.
Reading map
Section titled “Reading map”- 1: motivation (worktrees gave the substrate; how do you turn N branches into clean main?)
- 2: what integration means in a multi-agent context
- 3: Pattern 1, shared origin (the default)
- 4: Pattern 2, per-agent fork (strongest isolation)
- 5: Pattern 3, shared worktrees (fastest iteration)
- 6: choosing between patterns (decision tree)
- 7: the lead orchestrator’s role (7 responsibilities)
- 8: the semantic-conflict failure mode
- 9: how the lead catches semantic conflicts (three layers of defense)
- 10: the Clawless 2026-06-04 sprint, a real-world example
- 11: four worked examples (shared-origin 3-agent feature, per-agent fork external contributor, shared-worktrees solo sprint, when-to-abort)
- 12: OSS / microservice / long-term-feature-branch comparisons for experienced devs
- 13: a useful frame for managers and TPMs
- 14: L15-L16 foreshadowing
- 15: five team-scale scenarios
- 16: stay-calm psychology (the skill is learnable, the patterns are real)
- 17: closers and voice anchor
What this lesson deliberately does not cover
Section titled “What this lesson deliberately does not cover”- Specific multi-agent framework choices (LangGraph, AutoGen, custom orchestrators), those are tooling decisions, out of scope for a git workflow lesson
- AI agent capabilities or evaluation, adjacent topic, separate track
- Specific build/test toolchain choices, language-agnostic in this lesson
- Branch naming conventions for agent fleets, light convention suggestions in the worked examples; deep coverage out of scope
- Cost or rate-limit management for AI agents, operational topic, separate from git workflow
Estimated reading + practice time
Section titled “Estimated reading + practice time”70-90 minutes for the lesson body, plus 35-50 minutes for the practice drills. The lesson is the longest in Phase 4 because it’s where the most distinct new mental models live. Most learners spend the most time on the semantic-conflict catching section, it’s the part that doesn’t appear in any non-multi-agent git curriculum and is the highest-leverage skill.