Microsoft open-sources Agensh, a coordination harness for coding agents working without a manager

·1 views

Microsoft Research released Agensh as open source under the MIT license, along with its harness, experiment scripts, and a reproduction guide. The accompanying paper, posted to arXiv in September, describes up to 1,024 coding agents collaborating in one shared repository without a lead agent or central scheduler. Agents coordinate through a Gitea workspace, a Mattermost messaging layer, and a shared findings board that records dead ends and results. On hard reconstruction tasks, larger teams scored higher pass rates, though gains beyond small fan-outs were modest. A separate experiment found that one agent on a newer model outperformed the entire 1,024-agent fleet on the same task, highlighting that model choice can matter more than team size.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

UC Berkeley Math Professor Questions Value of Grading Homework Amid AI Use

A UC Berkeley math professor has concluded that grading homework no longer serves a clear purpose because of students' widespread use of artificial intelligence. The Daily Cal reports that the professor's concerns stem from the difficulty of telling whether submitted work reflects a student's own understanding. The article is based on the professor's comments to the student newspaper, and the story was shared on Hacker News, where it drew 15 points and 17 comments. The report centers on how AI tools are affecting traditional assessment methods in university mathematics courses. It does not detail any official university policy change in response.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Python 3.10 and Node 20 reach end of life: how to find pinned versions

Python 3.10 reaches end of life on October 31, 2026, and Node 20 already reached it on April 30, 2026. Ending support does not stop software from running, but security patches for new vulnerabilities will no longer be released. Because these runtime versions are often pinned in several files such as Dockerfiles, version files, and serverless configs, the article recommends auditing the whole codebase with a search. The author suggests choosing a runtime version with years of support remaining and notes Node 22 is supported until April 30, 2027. The piece also promotes a daily automated check that flags pinned versions 180 days before their end-of-life dates, while leaving the decision to upgrade to developers.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why lidar range falls short of datasheet figures: four conditions that matter

A lidar rated at 40 m may reach only 9 m outdoors, not because the device is faulty but because datasheet ranges are measured under unstated conditions. The article identifies four factors that usually explain the gap: the target's reflectance, which is often far below the 90 percent benchmark most specs assume; ambient background light, which can exceed a module's stated tolerance in direct sunlight; atmospheric visibility, which degrades returns in fog or rain; and integration time and update rate, which affect how long a sensor can collect signal. The author argues that range figures are only meaningful when paired with these conditions, and recommends testing with repeated measurements at several distances while recording the failure boundary.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Choosing Embedding Models for CRM-Linked Chatbot Retrieval

A DEV Community article argues that the best embedding model for a chatbot's RAG semantic search is the one that performs best on a labeled retrieval set while keeping the fields needed for CRM actions. Using an edtech sales-call assistant as the example, the author recommends measuring whether the right CRM policy, objection playbook, and follow-up rule are retrieved before weighing model size or cost. The proposed workflow defines a CRM action schema, builds test cases from real call summaries, compares candidate embedders under identical chunking and indexing, and validates generated actions separately. The piece includes a Python evaluator measuring recall at k and mean reciprocal rank, and suggests starting with 60 to 100 labeled queries and adding cases for every observed miss.

0 comments Read more at DEV Community