Part of Forge DevKit ecosystem

◇ forge-autopilot

Parallel agents without merge hell

The problem

Agent Teams = merge conflict chaos

"I literally can't merge their branches." Multiple agents editing the same files without coordination.

No conflict detection before execution

You discover file conflicts AFTER agents finish. Hours of work wasted on irreconcilable changes.

Autonomous agents with no strategy

Agent Teams is infrastructure. Without a strategy layer, agents work fast but break things faster.

forge-autopilot adds a strategy layer on top of Agent Teams: triage, conflict detection, and wave planning. Source: Claude Code docs

Backlog analyst classifies features by complexity, risk, and design needs. File dependency mapping.

Conflict detector identifies shared files between features. Classifies as SAFE, RISKY, or BLOCKING.

Groups conflict-free features into parallel waves. BLOCKING conflicts force sequential execution.

Wave 1: auth + payments (no overlap)

Wave 2: profile (depends on auth)

Teammates execute in isolated worktrees. Tech lead reviews. Auto-merge with conflict prevention.

Backlog→conflict detection→wave grouping→parallel execution. No merge conflicts by design.

Supervised (review each), semi-auto (review at checkpoints), full-auto (review at end).

Each teammate works in a separate git worktree. No file interference. Clean merges.

Features spanning multiple repos are detected and handled in triage phase.

Summary report after each wave: decisions, conflicts resolved, files touched, test results.

Dimension	Raw Claude Agent Teams	Forge DevKit
What it is	Infrastructure for spawning agents	Strategy layer: triage → conflict detection → wave planning
Conflict handling	None — discover conflicts after	Pre-execution conflict detection and wave grouping
Autonomy control	All or nothing	3 modes: supervised, semi-auto, full-auto

Architecture map powers conflict detection

Provisions isolated worktrees for each teammate

Syncs task status during batch execution