Shawn 'swyx' Wang, Smol.ai founder, details how Anthropic's Ultracode uses parallel subagents to execute complex developer workflows

@swyx yeah another framing I was debating is that it is a very effective way of adding (weak) determinism to non-deterministic knowledge work tasks

@swyx I use it quite a bit, it knows to select the right cheaper models for tasks, but sometimes needs a poke to remember.

I also set up the Workflow > Codex turducken, so Claude knows to route fiddly pedantic work to GPT 5.5.

@trq212 @swyx It's not even possible to experiment with the feature to learn its capabilities because it blows through usage limits so fast that people have to consider trading off their traditional Claude usage. Better not to touch it.

It's great. I use it at work where I'm not token limited.

But I tried to help my partner, a physician researcher with the $20 subscription, and it blew through her 5-hour usage limit in under 30 minutes.

I'm not sure what value there is in talking so much about this feature when the only people who can benefit are non-budget-constrained enterprises.

@swyx Have u seen what @sawyerhood is cooking w omegacode

@swyx Fanout is agent spawn plumbing at it's best!

@swyx

@trq212 @swyx with enough self-observation and self-remembering across jobs, sessions, memory, etc., a model could get close enough to determinism to be operationally useful. Not solved determinism, but enough continuity that behavior starts becoming steerable instead of merely prompted…

@swyx sounds like youre running a full dev team in ur terminal

whats the bottleneck when the repo isnt parallelized right?

@swyx i've had a decent time with it requesting efficiency in sub agents - was using it heavy with fable that way and it was cookin

@swyx the "set up your repo to parallelize properly" part is the bit everyone will skip and then wonder why fanout didn't help. the speedup is bounded by how cleanly the work decomposes, not the agent count. are you hand-partitioning the subagent tasks or letting it plan the split?

I've been obsessed with dynamic workflows since the day the feature dropped. Really pushes you to start thinking 'outside the box' of what's possible with subagent orchestration.

Throwback to running 101 Fable deep research agents in parallel last week😢 for a project I was working on. The results were superior and unlike any type of deep research I've ever done with other competitor models.

Eagerly awaiting the return of Fable...

@swyx It absolutely pumps through tokens on ultracode using fable high. But it was impressive to watch! Haha

@swyx Powerful underlying capability but I think it requires more precise control than they allow

@swyx interesting framing, subroutines but intelligent feels right

@swyx seems doable with codex with subagents as long as the plan and prompt is done right? I use this flow for multi worktree BC builds all the time

@swyx I like this idea, as long as you can intelligently vary effort levels by subagent, and really make sure you aren't using an agent on high effort, that could complete the task at the same level of proficiency, just faster and with less tokens, on low.

@swyx Would be interesting to see if this can be activated with kimi 2.7 running in claude code.

@swyx running ultracode orchestration with opus right now to fill the fable gap (not working). one model for architecture, one for implementation, one for review. token burn is real

@swyx A unified calling space for code and agent was the basis of https://github.com/zby/llm-do - but it never gained the traction I believed it should.