multi-agent factory — architect / coders / reviewer / tester
Architect (Opus 4.7)
plan + subtasks + interfaces
Coder A (Sonnet 4.7)
worktree A + Daytona
subtask: parser
Coder B
worktree B
subtask: cache
Coder C
worktree C
subtask: api
Coder D
worktree D
subtask: migration
merge coordinator
3-way merge + LLM conflict resolver
Reviewer (GPT-5.4)
cannot approve diffs it authored
Tester (Gemini 2.5 Pro)
run suite in clean sandbox
PR open
metrics
SWE-bench Pro pass@1
wall-clock vs single-agent
false-approval on bug probe
token amplification ratio
merge conflict rate
target:
pass@1 > single-agent
speedup 2-3x on parallel tasks
token amp < 2x
false-approve < 5%
SWE-AF / MetaGPT / AutoGen shape
handoff accounting
every role boundary = span
A2A-typed messages
file-backed task board
Langfuse role tag + token
post-mortem histograms:
plan vague / merge conflict /
reviewer false-approve /
tester flake
shipped with every failed run
handoffs are the failure surface