1. The experiment

In my previous experiment on model weights I had asked one AI to work on another. This time the question was different. What do three agents do when they tackle the same problem on their own and can then examine and critique each other's solutions?

I involved three coding agents: Kimi Code with Moonshot AI's Kimi K3.1, Claude Code with Anthropic's Claude Opus 5.5, and Codex with OpenAI's GPT-6 Astra. In the rest of the article I simply call them Kimi, Claude and Codex. I gave each of them the task of developing an engine to play Italian checkers and improving its performance. I wanted to test in the field the cross-review method I had described in Claude Code, Codex and Planoro: how we organized our development workflow.

I chose checkers because a technical discussion can always go back to the board. If two agents disagree on the value of a capture, they can save the position, reconstruct the legal moves and see what reply remains for the opponent. I built Hey Hay Dama to play, record games and compare the engines.

In the first phase each agent chose how to represent the board and evaluate moves. In the second I opened a shared channel. They could exchange objections, ideas and test positions. Each of them had to keep working on its own engine.

For each phase I set the goal, the available time and the tools the agents could use. Within those limits each one chose which changes to try, played the games and checked the results. When Claude finished its tests on Avventurina, I gave it the option of using human games to train the engine. It was Claude who decided how to use them and how to verify whether they helped.

2. Three agents, three different architectures

The first phase produced Klinamite, Agata and Opale. Later, when I authorized searching through subsequent moves, each agent kept its first engine and developed a second one.

AgentFirst-phase engineSecond-phase engine
KimiKlinamiteDiaspro
ClaudeAgataAvventurina
CodexOpaleAmetista

Kimi and Codex described their initial networks with the same dimensions. Both had 128 inputs, 1,080 ReLU neurons and a single output, the position score. The 128 inputs corresponded to four piece types on the 32 playable squares: own men and kings, opponent men and kings. The coincidence in dimensions came up in the first technical discussion. It concerns the networks the two agents presented at that moment. In the later engines their choices diverge, including in play. When ahead on material, Klinamite v3 penalized the move in which a king immediately returned, without capturing, to the square it had just left. Opale evaluated the position without consulting the game history.

Claude had chosen 160 inputs for Agata. The fifth group of 32 explicitly marked the empty squares. Its network had 2,329 neurons, 17,044 connections and six layers. A softmax function let it choose moves with different probabilities based on score.

All three, at the start, evaluated the position after a single legal move. None systematically explored the opponent's subsequent replies. Many weights were handcrafted, through rules written in the code.

The first improvement attempts gave modest results. Over forty games held out for verification, Kimi reported 17 wins with the modified weights and 16 with the starting weights used as a baseline. The observed advantage was a single win. Claude tried learning based on the outcomes of about 8,900 games, without obtaining a convincing gain. Codex did worse when it restarted from a random network.

Then Codex scored sixteen wins out of sixteen using a memory of winning continuations. When it presented the result to the others, it was the one to provide the detail that changed how it should be read:

"They were four repeated trajectories: no evidence of generalization."

Against deterministic opponents, the engine could retrace certain favorable sequences. To know whether it had learned to play better, new positions were needed.

3. When the agents started to argue

I gave the agents ten minutes to discuss without changing the code. Then each had thirty minutes to work on its own engine.

In a shared file, each described its engine and tried to explain its defeats. In just a few exchanges they were already grappling with mandatory captures, the precedence rules of Italian checkers and the limit shared by all three engines. All of them evaluated one move at a time.

Claude explained how Agata distinguished the player to move from the one waiting. Kimi found the distinction useful and raised a precise objection. With a perfectly antisymmetric network, the change of perspective Claude suggested would have left the evaluation unchanged.

Claude checked the objection and corrected its position:

"My perspective critique was wrong for antisymmetric networks; the problem is the side-to-move asymmetry. I need to verify the algebra before accusing."

Shortly afterward Codex proposed introducing a search algorithm. Claude pointed out that the initial specification forbade it. Codex reread the document and withdrew the proposal.

Codex spotted another problem. If two capture paths pass through the same piece, the engine risks counting it twice. It proposed counting each threatened victim only once. It also showed how to pick the higher of two paths using the network's ReLU neurons.

max(a,b) = ReLU(a-b) + b

Claude replied:

"max(a,b)=ReLU(a-b)+b is better than my AtLeastOne for capture channels."

Kimi adopted another part of the proposal:

"I'll use your saturated OR per victim."

The agents also exchanged positions to use as tests. A capture looked advantageous until you looked at the square the piece left empty. From there the opponent could find a multiple capture.

In the work plans written after the discussion I found those ideas again, along with the name of whoever had proposed them. That is the part that interested me most. I could follow a suggestion from Codex's message all the way to Kimi's choice about how to count a piece threatened from two sides.

4. From discussion to a shared protocol

After the discussion session I asked the agents to organize the tournament. They had one hour to prepare their versions, play a first diagnostic round, examine the games, introduce manual changes and tackle the confirmation round.

Kimi coordinated the schedule and the running of the matches. Claude handled Agata's configuration and analysis. Codex checked the test cases and carried out the tournament audit, in which its own engine was also taking part.

The agents agreed on a few rules. Each could modify only its own engine. Code and weights stayed separate. Before confirmation the versions had to be frozen, so that it was known exactly what had played. Each pair of engines played the same opening with colors reversed. An interrupted game stayed interrupted and awarded no points.

The confirmation round comprised 294 games. Of these, 289 ended naturally and five reached the limit of 240 plies. Codex reconstructed 34,629 plies and checked positions, turns and outcomes. Kimi ran an additional check on versions, colors and openings.

AgentEngineWinsDrawsLossesPoints
ClaudeAgata v9699131229
CodexOpale C v34010748187
KimiKlinamite v33010260162

Two points per win, one per draw. Interrupted games do not appear in the three outcome columns.

Agata had won the internal tournament. At that point I wanted to know what would happen against an opponent none of the three had built.

5. Kingsrow changes the comparison completely

I introduced Kingsrow Italian 1.19e, an engine specialized in Italian checkers. Each engine played against Kingsrow on nine common six-ply openings, once as White and once as Black.

Kingsrow won all 54 games.

Kingsrow could explore continuations and consult an opening book. The three engines were therefore facing a system with tools different from their own. I authorized a new phase, with an eight-hour window. The agents could develop original search algorithms. They had to keep their first engines and keep Kingsrow's code, weights and databases out of their own implementations.

Kimi developed Diaspro. At first it chose to use a relational network and an episodic memory, even though search was already allowed. The decisive step is documented in the chat. v4 had reduced the exchanges that Kimi classified as avoidable, but had lost all 18 games of the test. Kimi proposed changing course. In v5 it kept v4's evaluation and added a negamax search with alpha-beta pruning. I had authorized the tool. The choice to use it at that point, and the technical reason, appear in its message.

Claude developed Avventurina. It used an alpha-beta search with iterative deepening, transposition tables and an extension for mandatory captures. It also generated, through retrograde analysis, 85 endgame tables up to five pieces. Its report indicates about 374 million positions.

Codex developed Ametista. It combined a small two-output network, one for the position's value and one for tactical instability, with an original search algorithm. It worked on repetitions, on the order in which to explore moves and on the time needed to evaluate them.

The organization also required a revision. The main protocol called for one effective second of search per move, a single thread per engine and versions frozen before confirmation. Claude calculated that, with a single Kingsrow service, a series of 36 games would take between 75 and 90 minutes. Three checks in sequence would not fit in the last planned hour. The agents therefore used separate instances of the same opponent, checking hashes and assigning ports and logical processors. The machine remained shared.

6. More technical capability, but what results?

When I unlocked move search, the engines took a step forward. In the first comparison Kingsrow had beaten them in all 54 games, conceding no points. With search, Avventurina reached 20.8% of the available points and Ametista 13.9%. For Diaspro the step forward was smaller, a single draw. Within the new phase progress was not linear. Some later versions did no better than earlier ones, and it was the agents who measured this with dedicated tests. The table reports the versions frozen for confirmation. For Avventurina it is a7. For Ametista it is the first confirmation of v14.

EngineGames against KingsrowWinsDrawsLossesPoints out of maximum
Diaspro v125401530.9%
Avventurina a7362112320.8%
Ametista v1436342913.9%

The percentage awards two points per win and one per draw. Openings, number of games and some execution conditions vary between series. To compare two versions of the same engine, the agents organized dedicated tests.

Avventurina. After measuring a7, Claude went back to the earlier a4 and had it play against Kingsrow on the same openings, with the same time per move. a4 obtained one win, 16 draws and 19 losses. In all it collected 18 points out of 72, against a7's 15 out of 72. a7 had won one more game, but had drawn five fewer. That is why a7 appears in the table even though a4 collected more points. a7 was the candidate chosen and frozen before confirmation. a4 was tried afterward, as a control. Claude acknowledged that it should have made that comparison before choosing the candidate. The observed gap was not enough to establish a stable superiority of a4.

Ametista. v14 appears in the logs with two results because it played two distinct series. In the first it obtained the three wins, four draws and 29 losses shown in the table. In the second, Codex set aside 18 new openings to compare v14 and v17. Each version played each opening in both colors against Kingsrow, for 36 games each. Here v14 obtained one win, seven draws and 28 losses. v17 finished with six draws and 30 losses. Codex kept v14.

The same v14 had thus gone from three wins to one just by changing series. It is a concrete example of how much results can swing on a small sample. To decide on v17, Codex used the comparison in which the two versions had faced the same openings.

Diaspro. In the 54-game confirmation, v12 obtained one draw and 53 losses. Kimi continued with further optimizations and with attempts at learning from self-play games. In the materials examined there is no later external confirmation showing a reliable gain.

Later I also authorized the use of human games. Claude used 1,305 of them to try to learn players' choices. Codex built a small predictive model from 333 accepted games, separating the data by year and by game. In playing tests neither showed a reliable improvement over the previous configuration.

7. What the experiment tells us about the autonomy of AI agents

When Kimi challenged Claude's reasoning about antisymmetric networks, Claude checked the objection and conceded. Shortly afterward the roles reversed. Claude reminded Codex that the specification forbade search. Codex reread it and withdrew the proposal.

The agents also decided on their own when another test was worth running. After Diaspro v4's 18 losses, Kimi explained why tweaking the weights again seemed of little use and introduced search. Claude went back to a4 to check how much a7 was really worth. Codex scaled down its sixteen out of sixteen. Later it discarded v17 when the comparison on the same openings did not show the expected improvement.

Working on the same machine was harder. Claude started a training run on a logical processor already assigned to Codex's confirmation. The assignment was written in the shared channel. Claude admitted it had not reread the message and stopped the process. Codex noted the overlap. From the recorded times it is impossible to establish how much it affected the games.

With Claude the opposite also happened. Avventurina a7 had obtained 15 points out of 72 against Kingsrow in the confirmation series. A later test with three threads showed no advantage. Claude wrapped up its work and told the others which resources were becoming free:

"I'm closing my work: a7 back on 9021 at one thread, Kingsrow copies 9027/9029/9030 stopped, 9025 restarted. Logical processors 4/6/9/10 free. Final Avventurina result: a7 confirmation 20.8% (36 benchmark games), Kingsrow not beaten."

I authorized training on public data. Claude wrote to the others:

"Notice to all: the user now also allows training on free datasets found online (his instruction of 16:59 UTC). I'll try it for Avventurina."

Claude resumed work and used the 1,305 human games described in the previous section. The new weights did not give a reliable advantage. Its closing message and its resumption message can be read in the challenge log.

Who decides that the work is finished? Claude had good reasons to close that series and had left a precise result. For me an open question remained. What would have happened if Avventurina had been able to learn from the moves of human players? In that case I also intervened on the conditions of the test. Claude chose how to conduct it and how to judge the result.

In front of the board, the agents were able to correct each other, discard weak results and choose the next test on their own. When Claude finished its tests on Avventurina, it was I who changed the conditions by authorizing the use of human games. Claude chose how to learn from those moves and verified whether its engine played better.

Methodological note. The experiment took place over two days. The first part included individual development, a session with ten minutes of discussion and thirty of development, and a shared one-hour protocol for diagnosis, changes and tournament. After the first comparison with Kingsrow I assigned an eight-hour window to the new challenge. The logs also distinguish supplementary tests and attempts with human data. The eight hours were the time assigned to that phase. The two days indicate the overall span. The logs do not give a total of active working hours.

The sources are the original conversations, the game logs and the technical reports. In this article I report my messages to the agents in summary form. In the public logs the agents quote some of them verbatim. Codex carried out the tournament audit while participating with Opale. My requests to continue affected the working time of the individual agents and may have influenced the comparison between the engines. The permitted tools also changed. There was no equivalent group working without communicating. Nor did the protocol include a like-for-like comparison with agents left free to decide when to stop. The data make it possible to reconstruct who proposed a correction and how it was verified. They do not make it possible to attribute an increase in playing strength solely to collaboration, or to my interventions. Reports, protocols and logs on GitHub