Post

Sam Minnée
@sminn.ee

Anthropic have put their toe into agent orchestration. Something missing from this, and I gas town too, is input multiplexing: if I’m running a lot of agents and I’m not a cowboy, there are a lot of steps to approve and outputs to review. Bouncing between Claude windows is confusing & tedious.

Andrew Lilley Brinker@alilleybrinker.com

"Move Over Gas Town, Claude Has First-Party Agent Orchestration" Anthropic has announced an experimental setup for orchestrating multiple agents, so I do a quick comparison between that, Gas Town, and Claude subagents.

alilleybrinker.comMove Over Gas Town, Claude Has First-Party Agent Orchestration — Andrew Lilley BrinkerAnthropic has announced an experimental system for first-party multi-agent orchestration, called 'Agent Teams'.

0 reposts3 likes

Sam Minnée
@sminn.ee

I’d be interested a multiplexed view of this, where all the input requests were shown in one place, and perhaps grouped into small requests (“run this kubectl command?”), plans to review, and code to review.

Sam Minnée
@sminn.ee

Plans & code arguably deserve dedicated review interfaces, but maybe a multiplexed agent oversight UI could have optimised tooling for that work. Big picture, an agent-first alternative to GitHub.

Sam Minnée
@sminn.ee

Compared to the fully-remote agents, I still like the ability to dive into their thought-streams and in-progress codebase from time to time, but a better high-level view is needed as well.