It talks about how they developed multi-agent research systems -
1. ๐๐ฒ๐๐ผ๐ป๐ฑ ๐ฆ๐ถ๐ป๐ด๐น๐ฒ ๐๐ด๐ฒ๐ป๐๐
โ Single agents hit context and time walls on complex research tasks, where bigger models aren't the solutionโorchestrated coordination is
โ When one Claude model stalls on broad queries like "list every S&P 500 IT board member," parallel subagents with separate context windows can explore different angles simultaneously
2. ๐ง๐ผ๐ธ๐ฒ๐ป ๐๐ฐ๐ผ๐ป๐ผ๐บ๐ถ๐ฐ๐
โ Multi-agent systems consume ~15ร more tokens than standard chat, but token volume alone explained 80% of performance variance on BrowseComp evaluations
โ The trade-off is clear: spend more tokens to unlock capabilities that single agents fundamentally cannot achieve
3. ๐ข๐ฟ๐ฐ๐ต๐ฒ๐๐๐ฟ๐ฎ๐๐ผ๐ฟ ๐ฃ๐ฎ๐๐๐ฒ๐ฟ๐ป
โ A lead agent analyzes queries, develops strategy, and spawns specialized subagents with focused objectives
โ Each subagent operates independently with its own tools and context window, then returns distilled insights, preventing chaos of unstructured multi-agent interactions
4. ๐๐ป๐๐ฒ๐น๐น๐ถ๐ด๐ฒ๐ป๐ ๐๐ผ๐บ๐ฝ๐ฟ๐ฒ๐๐๐ถ๐ผ๐ป
โ Traditional RAG uses static chunk retrieval while multi-agent research uses adaptive, multi-step search
โ Subagents act as intelligent filtersโexploring different aspects in parallel and compressing vast information into key insights for the lead agent
5. ๐๐ผ๐ผ๐ฟ๐ฑ๐ถ๐ป๐ฎ๐๐ถ๐ผ๐ป ๐๐ผ๐บ๐ฝ๐น๐ฒ๐ ๐ถ๐๐
โ Coordination complexity grows rapidly, with early agents making errors like spawning 50 subagents for simple queries or scouring endlessly for nonexistent sources
โ Success requires embedding clear heuristics: delegate with explicit objectives and output formats, scale effort by complexity (1 agent for facts, 10+ for open research), and start broad then narrow based on findings
6. ๐ฃ๐ฟ๐ผ๐ฑ๐๐ฐ๐๐ถ๐ผ๐ป ๐ฅ๐ฒ๐น๐ถ๐ฎ๐ฏ๐ถ๐น๐ถ๐๐
โ Agents are stateful and non-deterministicโsmall changes cascade into large behavioral shifts that can't be fixed with simple restarts
โ Solutions include checkpointing progress, full execution tracing, rainbow deployments for live updates, and robust retry logic
7. ๐ข๐๐๐ฐ๐ผ๐บ๐ฒ ๐๐๐ฎ๐น๐๐ฎ๐๐ถ๐ผ๐ป
โ Unlike traditional software with fixed paths, agents take different valid routes to the same goal, making standard evaluation methods inadequate
โ Effective evaluation uses LLM judges for scalable assessment plus human reviewers for edge cases, focusing on whether agents achieved the right outcomes through reasonable processes
8. ๐๐๐๐ป๐ฐ๐ต๐ฟ๐ผ๐ป๐ผ๐๐ ๐๐๐๐๐ฟ๐ฒ
โ Current systems run subagents synchronouslyโthe lead waits for all to complete, which simplifies coordination but limits parallelism
โ Asynchronous execution where agents spawn new agents mid-task promises major performance gains, but introduces challenges in state consistency and result merging
P.S. Check out my profile for more resources on AI Agents ๐
This post was originally shared by on Linkedin.