In my recent experience it seems to be a contextual issue. Llms are constantly being dropped into new situations where they have to rediscover high level and cross cutting information about the system they're working on from contextual clues.
The internal representations of this state and its projection back out to human language wouldn't be as concise as that of a practitioner or team that develop their own verbiage and ontology over time molded to their system.
This verbosity might get better as we figure out better ways for agents to learn long term and use that knowledge to adapt to the users and projects over time.
There might also be some good harness improvements we could consider like forked output streams or multiple long lived filter subagents to ensure that output appropriate for thinking is separate from code output and separate from output given to the user driving the session.
The internal representations of this state and its projection back out to human language wouldn't be as concise as that of a practitioner or team that develop their own verbiage and ontology over time molded to their system.
This verbosity might get better as we figure out better ways for agents to learn long term and use that knowledge to adapt to the users and projects over time.
There might also be some good harness improvements we could consider like forked output streams or multiple long lived filter subagents to ensure that output appropriate for thinking is separate from code output and separate from output given to the user driving the session.