Co-founder of agent-swarm here. My co-founder Eze wrote this article today on an interesting topic: how making our agents build software to automate themselves saved us +$17k in the last month in token prices.
This is something we've been doing for a while, pushing for code-mode, repeatable scripts, workflows.
This also allowed us to build a dreaming workflows that does 20 more steps (more reliably) than the previous prompt based one and even reduce the cost by 6% (not much but still).
I'm a big fan of moving mechanical process from agent to script (determinism). Things like collecting and rendering prior sessions to markdown or around the PR process (have a post nearly done on this one myself).
Why have agents call APIs directly, or even write the code per-session, versus having them write a script that human or machine can use?
It is still unclear to me what people mean by code-mode. My agents regularly write inline code in-session with bash, python, node all the time, on their own without any prompting. I'd prefer they put these things in permanent scripts we can reuse or rework.
5 comments
[ 0.22 ms ] story [ 21.4 ms ] threadCo-founder of agent-swarm here. My co-founder Eze wrote this article today on an interesting topic: how making our agents build software to automate themselves saved us +$17k in the last month in token prices.
This is something we've been doing for a while, pushing for code-mode, repeatable scripts, workflows.
This also allowed us to build a dreaming workflows that does 20 more steps (more reliably) than the previous prompt based one and even reduce the cost by 6% (not much but still).
Cheers,
Why have agents call APIs directly, or even write the code per-session, versus having them write a script that human or machine can use?
It is still unclear to me what people mean by code-mode. My agents regularly write inline code in-session with bash, python, node all the time, on their own without any prompting. I'd prefer they put these things in permanent scripts we can reuse or rework.
I'm more fond of a world where many many people build personal harnesses on SDKs. More experimentation is good this early on.