Manager-less DreamTeam swarm wrote trustcheck for $5.68 and passed 13 of 13 hidden tests

DreamTeam is an open-source Python framework (MIT license) for a self-organizing swarm of AI agents with no manager agent. The agents get a single goal, a shared task board, a shared git folder and the ability to message each other, and they work out the rest themselves. The author tested it on four tasks with independently verifiable answers: three in mathematics and one in engineering.

The most telling is the engineering one: a swarm of 8 agents running the cheap deepseek-v4.1-flash model wrote trustcheck, a tool that answers the question "can these tests be trusted?", in about three hours for $5.

68. It was checked by a hidden Claude benchmark: 13 git repositories with traps and a known correct answer. An empty stub in place of the tool passes 1 of 13; the finished trustcheck passes 13 of 13.

Claude connected to the swarm over MCP as a user and external verifier.

A board instead of a manager

A session starts with a single task: "plan." The first free agent claims it atomically, so two agents cannot get the same task. The planner lays out the work: independent tasks first (literature, environment, data, trial experiments) so the others are not left idle, then dependent ones. Every task has a success criterion, an assignee, a reviewer and dependencies. An assignee cannot accept their own work: another participant does the review, and if there is none, the task goes to a human. If the swarm is short-handed, agents hire colleagues from a role catalog, within a limit.

When an agent lacks a tool, it writes one itself as an MCP server, and the whole swarm can use it. The swarm itself is also an MCP server, so Claude Code can observe it, control it or join the team as one more participant.

Four runs: from third place to an under-converged table

In the Erdős minimum overlap problem (the smaller the constant estimate, the better), 4 agents spent 1 hour 40 minutes and $0.51 to reach an estimate of at most 0.38090966633824946 on a step function with 800 steps. That beats the results of Haugland (0.380927) and AlphaEvolve (0.380923), but trails Hill Sampling (0.38088232) and TTT-Discover (0.380876). It is not a record, but it is third place.

In the selection amplifier problem, the goal is a graph on which a beneficial mutation fixates more often than in a well-mixed population, under both Moran process rules at once, Birth-death and death-Birth. 8 agents spent 2 hours 40 minutes and $3.97 to find a strict simultaneous amplifier on 14 vertices, whereas the previously known example had about a thousand.

The effect is tiny: amplification for r from 1 to about 1.016 and a gain of no more than ~10⁻⁵. Claude certified the result at four values of r, and showed by exhaustive search that no unweighted graph of this kind exists for N ≤ 8.

The author checked novelty against only a few 2024–2026 papers and does not call the result "first."

The most instructive run was the Thomson problem in a disk: place N identical charges so that the repulsion energy is minimal. In a little over four hours and for $3.98, the swarm produced configurations for all 131 values of N from 20 to 150 and reproduced four recent records with three independent methods.

On 57 values of N, its energy came out lower than the Amore-Zarate table by 10⁻⁹…10⁻⁶. But Claude polished the original configurations with Newton's method and got the same energies: the published points were simply under-converged.

In the end, on 128 of 131 values of N the minimum matched the table; on N = 131–133 the swarm's minimum is higher by 0.07–0.08, and none is lower.

The author's conclusion: the swarm readily declares success based on a raw difference in numbers, and review inside it checks that the task was done, but not that the result is correct. Only verification from outside protects against this.

In the engineering run, trustcheck performs mutation testing within the diff, issues verdicts on individual tests and looks for tampering in commit history. The final package contains 231 passing tests. The v2 and v3 trap sets were added along the way: at one point a fix for the old set caused a regression, and only the new set caught it. After the second round ($4.15), on the fresh v4 set the stub scored 2/12 and the tool 12/12.

What remains unsolved

On DreamTeam's own ~1400 tests, trustcheck takes 13–25 minutes, and its score on this code is 0.60 against a target of 0.7. In an A/B test on three historical cookiecutter bugs, the trustcheck skill did not help: the hidden tests passed 29/29 in both arms, and wrong fixes were caught 13/13 without the skill and 12/13 with it, at a cost of $1.00 versus $0.81. Both groups hit the ceiling. Along the way it emerged that an agent with network access can download a version of a library in which the bug is already fixed, which is why a no-internet session mode was added.

The runs exposed other problems, which the author turned into fixes: the planner made the final report depend on optional tasks, the architect spent 30 minutes unable to escape reasoning truncated by the output limit, and when asked to "wrap up," the agents launched one more "quick" computation on 16 processes. There is now a wrap-up mode with no new background processes, and the output limit doubles after truncation. The practical takeaway for anyone building their own agents: useful feedback comes from long runs on real tasks, not from short demos, and you always need an independent verifier.