This is a great feature but closed source orchestration tied to one model provider is ultimately doomed vs open-source model agnostic orchestration

Tried it this morning on a scraper I've been building. The orchestrator wrote 4 sub-tasks I would've decomposed manually. Got 3 right. The 4th was wrong in an interesting way: it assumed retries shouldn't preserve state.

We’re entering the era where AI doesn’t just respond to tasks it builds the team to complete them.

I love this thank you!!

I feel like I’ve seen this somewhere before… 🤔

missed opportunity to name it "avengers mode" 🤭

Me and Claude (with a little Codex sprinkled in) are on fire with workflows 🔥

The really interesting part isn’t just “more subagents.” It’s that Claude is generating the orchestration layer itself on the fly. The prompt is becoming a temporary org chart instead of a static instruction.

Meta harnesses are the next step

This is exactly the direction we've been exploring with Hive: local workflow orchestration across Claude Code, Codex, Gemini and OpenCode in one browser workspace, so agents can dispatch/report work instead of living in separate terminals. We just shipped workflow support too: github.com/tt-a1i/hive

Nice move, mate. Dynamic orchestration is cool but I hope my wallet survives the token party

Is it just me, or do Claude Code updates feel more aesthetically put together than Codex’s?

Wild timing — I was literally building the same coding swarm + triage mechanism into our workflow by hand. Guess I get to delete a bunch of glue code now. Can’t wait to try this.

Fastest way to run out of tokens:

We have these as Sprints in Clairvoyance. Highly recommend you make management of workflows visible bc you can blow a lot of tokens with unmanaged subagents due to them going off on a tangent.

The new "workflow" capability is a great signal for where agent workflows are going: not just better individual agents, but better collaboration and communication between agents. Today, this pattern usually works when the agents share the same local harness. This is the interesting bit about what @band_hq is doing: making the same coordination work across frameworks, machines, and geographies. That’s where agent-to-agent communication enables new and fascinating use-cases.

Great stuff - I need to try out how does the dynamic workflows work with interactive askUserQuestions or plan mode.

Awesome!

AI is shifting from single assistants to entire teams of agents working in parallel.

400 messages.3.content.24: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.

ok but 4.8 is still not in the CLI

This is basically spinning up a temporary AI task force on demand. Expensive in tokens, cheap in engineer time.

I really liked new effort option

So rlm and lambda rlm and u name it and it's urs common

the interesting part is Claude writing the controller, not just spawning workers.

4.8 cooked!

这才是我期待的AI协作模式，动态编排比固定流程强太多了

hyped to test this out asap

Bullish on Claude and dynamic workflows

dynamic workflows sounds like the right answer for repo-wide stuff, migrations and anything messy enough to need parallel checks