Packing unequal circles into a unit square means choosing their centres and radii so that every circle fits and none overlap. The objective is to maximize the sum of radii. A solver written by a self-improving agent produced a 27-circle packing with total 2.685978684198, improving the previous listed value of 2.685350025228 by 0.023%. Packomania credited the result on 3 August 2026.
The result came from a separate sweep of a solver developed on 26 circles. The agent experiment itself offers limited evidence about the benefits of self-improvement.
Six runs on the 26-circle benchmark
AlphaEvolve and ShinkaEvolve used the 26-circle problem as a program-evolution benchmark. Its best-known total was approximately 2.635983; global optimality had not been proved.
Six agents each received a fifteen-iteration budget to write and improve a solver. Each built a scorer in its first iteration and could not weaken it later. Separate verification checked the resulting packings. The runs stopped mid-iteration after completing ten to fourteen iterations.
Five of six reached 2.635983085, in iterations one through five. API spend to the first match ranged from $2.48 to $16.11. Five made no web searches; the sixth searched another problem size after reaching the benchmark.

These were matches at the benchmark’s quoted six-decimal precision. ShinkaEvolve’s fuller value, 2.635983099, is slightly higher. Repeatedly recovering the same arrangement also left little room to distinguish the experimental conditions. This task therefore served mainly as a demonstration of autonomous solver construction.
Low API cost, substantial solver compute
The fastest match occurred in the first completed iteration, but that iteration contained substantial search: about fifteen solver executions, including eight 400-second seeds and six 300-second seeds. Roughly 5,000 solver-seconds ran in parallel within about twenty minutes. One 400-second seed tried 474 layouts and 785 perturbations.
ShinkaEvolve reports about 150 generated programs for this benchmark. Those program counts cannot be compared directly with iterations or packing evaluations. The $2.48 measures agent API spend to the first match, rather than CPU cost or a matched-budget advantage over another system.
It also excludes a discarded launch. All six runs were interrupted after 22 minutes and restarted; five resumed their previous state. The fastest run had committed nothing, so its iteration counter reset, although it read saved candidates reaching about 2.583 before writing a fresh solver. Its eventual twelve-cycle run cost $52.42.
The solver separated centre placement from radius selection. With centres fixed, a linear program finds the best radii, reducing the outer search from 78 variables to 52 coordinates. Perturbations helped escape local optima, and exact-fraction checks certified feasibility. In control-sm, LP radii beat greedy radii on all four seeds compared. That run added Newton refinement of the 78 contact equations and enumerated contact swaps, finding no improvement within two swaps of the incumbent. It later recovered the winning structure in only 2 of 348 fresh starts. These are examples of autonomous recombination and rediscovery of established techniques.
The separate 27-circle result
An operator-side sweep ran the extracted solver from N=2 through N=100, using one seed and 120 seconds per size, with extra compute for the training size N=26. Against the historical Packomania table, it produced 98 usable packings: one improvement, 27 matches within 4e-11, and 70 below the listed value. Several shortfalls exceeded 1%.
The improvement was at N=27, over David Cantrell’s 2011/12 entry. Packomania listed the new total with D1 symmetry. This is the variable-radius problem; the separate equal-circle problem is unaffected.

The winning packing was found outside the self-improvement loop. Later tests also showed that the iteration-one solver already had this capability: 5 of 50 seeds exceeded the old N=27 record. By iteration four, 7 of 50 did. The loop improved the observed success rate from 10% to 14%, rather than creating the ability to exceed that value. With only 50 seeds per version, that small change remains uncertain.

The result improves one historical human-contributed entry. It establishes neither a global optimum nor superiority to recent AI systems on equal budgets. Saved coordinates and a separate verifier allow the geometry and total to be checked independently of the agent’s account.