Insights, news and research from the team building Cominty
What we think enterprises should do about agents, what we have shipped, and what we measured. Longer than a changelog, shorter than a paper.
Agentic AI becomes part of everyday work: observations from 500 professionals
What Cominty observed among 500 users of its agentic portal across 10 client companies between July 2025 and July 2026: longer work sessions, six times more end-to-end delegation, and outputs grounded in company context.
- 17 September 2026[insights]
Agentic Harnesses: From an LLM Loop to a production Agent
An AI model produces outputs. An agent pursues an objective, may use tools, and maintains the state required for its execution. The agentic harness is the layer that turns this behavior into a controlled system.
Fariha
- 13 September 2026[engineering]
Does enforcing JSON make answers worse?
Chat products often wrap the reply in a JSON envelope to get related queries or citations in one call. We tested six ways to do it on nine models and five benchmarks: reasoning and code inside JSON mostly survive, instruction following and the look of the writing pay, and one sentence in the schema moved Gemini 2.5 Pro from 0% to 87%.
Cominty Engineering
- 11 September 2026[engineering]
Can an LLM feel the road? Fast, cheap drive-time estimates across France
While tuning an agent that plans field rounds across mainland France, we needed a fast, cheap answer to “how long is this drive?” We compared a distance-only curve, a residual neural network, and a language model guessing from memory: the network cuts error from 13.70% to 9.53% MAPE, and the language model falls apart under 20 km.
Cominty Engineering