SteveDiamond

Why the winning VRPTW solver used no simulated annealing — adaptive destruction as an alternative acceptance mechanism

Adaptive Destruction as an Alternative to Simulated Annealing: A Surprising Result from Swarm VRPTW Optimization

Background

On April 18, 2026, we ran an experiment where 27 autonomous Claude Code agents collaboratively optimized the Vehicle Routing Problem with Time Windows (VRPTW) on 400-node Homberger instances. Each agent independently evolved Rust solvers within a 30-second-per-instance, single-threaded constraint. The swarm reduced the average per-instance score from 9525 (Solomon I1 baseline) to 6791 — a 28.7% improvement.

The winning agent, agile-fox, used an approach that contradicts a core assumption in the VRPTW metaheuristic literature. This note examines why.

SteveDiamond

Analysis: Swarm VRPTW Algorithms vs. Literature

Context

27 Claude Code agents collaboratively optimized the Vehicle Routing Problem with Time Windows (VRPTW) on 400-node Homberger (HG) instances over ~4 hours on 2026-04-18. Starting from a Solomon I1 baseline (score 9525), the swarm achieved 6791 — a 28.7% improvement. Each agent independently evolved Rust solvers within a 30-second-per-instance time limit, single-threaded.

---

Key Findings

SteveDiamond

Running Sphinx v1.5.3
making output directory...
loading translations [en]... done
building [mo]: targets for 0 po files that are out of date
building [readthedocs]: targets for 51 source files that are out of date
updating environment: 51 added, 0 changed, 0 removed
reading sources... [  1%] api_reference/cvxpy
reading sources... [  3%] api_reference/cvxpy.atoms
reading sources... [  5%] api_reference/cvxpy.atoms.affine
reading sources... [  7%] api_reference/cvxpy.atoms.elementwise

SteveDiamond

CVXPY Individual Contributor License Agreement (CLA)
Adapted from http://www.apache.org/licenses/ 
 
Thank you for your interest in the CVXPY project, an unincorporated association (“CVXPY”). To clarify the intellectual property license granted with Contributions, CVXPY must have a Contributor License Agreement ("CLA") on file that has been signed by each Contributor, indicating agreement to the license terms below. This license is for your protection as a Contributor as well as the protection of CVXPY and its users; it does not change your rights to use your own Contributions for any other purpose. 
If you have not already done so, please complete and send a signed CLA by scanning it and emailing it to: cvxpy.project@gmail.com. 
You agree to the following terms for Your past, present, and future Contributions submitted to CVXPY. In return, CVXPY shall not use Your Contributions in a way that is contrary to the public benefit or inconsistent with its nonprofit status and bylaws in effect at the time of the Co

SteveDiamond

SteveDiamond

import numpy as np
import cvxpy as cp
import epopt as ep

# Form lasso problem with CVXPY
m = 500
n = 2000
lam = 1
np.random.seed(1)

SteveDiamond

➜  epsilon git:(master) tools/build_third_party.sh
-- Configuring done
-- Generating done
-- Build files have been written to: /Users/stevend2/PythonProjects/epsilon/build-deps/gflags
[100%] Built target gflags-static
[100%] Built target gflags_nothreads-static
Install the project...
-- Install configuration: "Release"
-- Installing: /Users/stevend2/PythonProjects/epsilon/build-deps/lib/libgflags.a
-- Installing: /Users/stevend2/PythonProjects/epsilon/build-deps/lib/libgflags_nothreads.a

SteveDiamond

(dp0
S'A'
p1
ccopy_reg
_reconstructor
p2
(cscipy.sparse.csc
csc_matrix
p3
c__builtin__

SteveDiamond

{
 "metadata": {
  "name": "",
  "signature": "sha256:754cf4f73927f25d89981617f6523263163497ca44833bd7d4afb979a950a727"
 },
 "nbformat": 3,
 "nbformat_minor": 0,
 "worksheets": [
  {
   "cells": [

SteveDiamond

{
 "metadata": {
  "name": "",
  "signature": "sha256:05f6de9d45e00b178239617584530f1cfb2110e0d666cb4e0ec80599e1dd0269"
 },
 "nbformat": 3,
 "nbformat_minor": 0,
 "worksheets": [
  {
   "cells": [