Microsoft Research tested a coding system that lets AI agents divide work among themselves rather than wait for a central coordinator. In a six-hour experiment, its largest team reproduced more of one software program's behavior than smaller teams did. The result shows that coordination can work at unusual scale, although the researchers did not test whether running so many agents is cost-effective.
The system, called Agensh, gives each worker access to a shared code repository, a messaging channel and a record of other workers' findings. Agents claim tasks, implement changes, check them and merge verified contributions. The researchers used Git-based branches to keep concurrent edits separate and a shared record to flag overlapping claims; they did not train a special manager to assign every task.
More agents passed more tests in a timed rebuild
To measure the approach, the team chose five difficult ProgramBench tasks. Each asks agents to rebuild the behavior of a reference program without internet access, then scores the result against tests. All runs used the same underlying model and a six-hour deadline. Across those five tasks, the mean share of tests passed rose from 19.31% with one agent to 28.78% with 128. That is a gain in the benchmark's test pass rate, not a measure of how much of an application would be ready for users.
The 1,024-agent configuration was tested on only one of those programs, the document converter pandoc. It passed 55.06% of tests, compared with 50.94% for 128 agents and 33.89% for one. The researchers spread the largest team across 16 computing nodes and staggered worker launches to reduce contention. Those conditions matter: the extra 896 agents improved the pandoc score by 4.12 percentage points over the 128-agent run, a much smaller gain than the jump from one to 128.
Cost and broader reliability remain open
The paper reports test results and examples of workers coordinating reviews and merges, but does not report the total model-token or computing cost of the configurations. It also does not show a 1,024-agent result for the other four programs or for routine software maintenance. For teams considering this design, the next useful comparison is whether the higher test score justifies the added resources across a wider set of tasks.
