avi
@avichalp.eth
enjoying reading these agent architecture blog posts recently in summary how claude deep research product work is as follows. there is a lead agent powered by opus 4. it accepts user's query and creates a plan. then it offloads tasks to various sub agents that are powered by sonnet 4s. sub agents mostly run in parallel and provide separation of concerns. sub agents themselves are capable of planning too. they often revise their plans and change trajectories if needed. tool usage and planning are interleaved during the whole process. secret sauce of a multi agent research tool is compression. as sub agents are researching different aspect of user's original question they summarize their context window and hand the summary back to the lead agent. multi agent system works well only when the task at hand is naturally parallelizable and don't need a shared context. most tasks are not. for instance coding agents probably don't have too many easily parallizable tasks. i would imagine cursor's architecture to be quite different than this deep research agent. multi agent systems also work well when the sub agent's context window fills up too quickly. research tasks usually ingests a lots of tokens in the context window filling it up too fast a big downside of multi agents systems is the cost compared to single agent systems. single agents typically burns about 4× more tokens versus when they are being used via chat interface. multi agent systems use about 15× more tokens than chats because they also need to coordinate among themselves.
2 replies
3 recasts
25 reactions
rubinovitz
@rubinovitz
Lmk if you find any other good ones!
2 replies
0 recast
1 reaction
rubinovitz
@rubinovitz
Also omg this was sent through your client?!
0 reply
0 recast
0 reaction
avi
@avichalp.eth
🫡
0 reply
0 recast
1 reaction