no major lab is working on that<\/strong>. Their communication style is carved in the underlying LLM weights. It can adapt to an interlocutor in the course of a conversation, but every new session resets that. They certainly don’t adapt to a whole team because they are usually tied to an individual’s machine\/acount.<\/p>\n\n\n\nWhy is that not dicussed more ? My guess is that those notions are somewhat fuzzy and difficult to define and evaluate. And companies focus on benchmarks (so, easyly-measurable metrics) that usually involve solving problems alone, not in collaboration with a team.<\/p>\n\n\n\n
An amnesic genius <\/h4>\n\n\n\n
The memory of coding agents is un-intuitively different from human memory. Their harness can help them remember word-for-word a conversation you had with them 3 months ago, but it usually doesn’t fetch that conversation’s memories by itself, or not in a reliable way. Currently a lot of work is focusing on precisely gathering the appropriate context for the task, but the focus is on code rather than past conversations.<\/p>\n\n\n\n
As a result, coding agents tend to follow the instructions given immediately before, but without taking into account the rest of the code properly. They’re like a coding genius, able to one-shot build simple applications better than a human could, but an amnesic genius that forgot everything about you every time you start a new conversation. They stick to neither your working style nor your coding style. They try to execute your instructions in one shot but often break existing code in the process. The code they add is often not maintainable and ends up adding more work for developers in the long term.<\/p>\n\n\n\n
The new session trick<\/h4>\n\n\n\n
A frustrating way in which coding agents’ memory is not intuitive, is how their memory resets between sessions but that’s sometimes an advantage. Sometimes, I spent hours on an issue with a coding agent without solving it; but after restarting a fresh conversation, the agent solved the problem immediately.<\/p>\n\n\n\n
This happens because of Context momentum<\/em>: in long chats, the LLM can get stuck in wrong assumptions made at the start, and doesn’t question them anymore. You can then spend hours trying to find an error where it isn’t, and solve it immediately after starting a new session, because the wrong assumption is not in the agent’s context anymore.<\/p>\n\n\n\nChaotic-Neutral compliance<\/h4>\n\n\n\n
LLMs don’t have feelings for you, the state of your project, or anything. They are trained to complete sentences, and not just any sentences: the instruction-tuning phase trains them to complete conversations between assistants and users. Therefore, they are very compliant.<\/p>\n\n\n\n
That is, they’ll do exactly what you said even if it’s so stupid that what you said<\/em> is clearly not what you want<\/em> (things like rm -rf \/<\/code> and rm -rf \/.<\/code>), and they usually won’t warn you. Because their answers are not deterministic and because their training is not perfect, LLMs have an unequal performance: sometimes they are impressive, sometimes they do something stupid, but always trying to follow your instructions.<\/p>\n\n\n\nThat makes them unreliable to some degree. Many developers will use an AI that can reliably implement ideas of a certain complexity rather than one that can execute on much more difficult tasks but sometimes fails in an unforeseeable manner.<\/p>\n\n\n\n
<\/p>\n\n\n