The trap hiding inside multi-agent systems

[00:00:00]Wait, your multi-agent system might only need one agent. There are three ways to set up agents: single, multi, deep.

[00:00:08]Single agent, one agent does the whole job. Reviewing your code for duplicate code, security holes, and messy formatting. Multi-agent, separate agents, each of its own job.

[00:00:20]Independent jobs, run them in parallel, and they combine the results. But, when the work is one connected job, and passing it from agent to agent, it will lose the details [music] at each step. So, when the work is one really connected sequence, the single agent is usually better fit. Stanford tested this by giving a series of agents and single agent with the same compute.

[00:00:43]Got the same or better results, and was cheaper to run.

[00:00:46]And the last, deep agent. One agent is in charge. It plans the job, sends sub-agents to check different parts of the code, and keep the track of what they find. Send this whoever is building the multi-agent system. Full breakdown is on the newsletter. Comment agent for the link.