Hermes made this quick explainer short on the new subagent controls your orchestrator Hermes Agent can use in v0.21.0 (the voice could be better but eh)

You can make videos like this too, with the built in manim-video skill in Hermes btw :)

Can the orchestrator Hermes Agent cancel or reprompt subagents dynamically with the new controls in v0.21.0?

Thats exactly what the video shows... cmon lol

@cystalphantom Do we need to enable anything or the orchestrator will know to monitor and steer subagents already with the new update?

@Teknium Does Hermes Agent require an explicit setting to steer and reprompt subagents, or does v0.21.0 handle orchestration automatically?

My favorite so far is steer. Excellent job with that. Really allows more fluid workflows.

this is the direction I find most interesting with agents, not making one model do everything, but giving it the tools to coordinate

these are two different measurements. one long session fills one context window, 1390 subagents means 1390 separate windows and the orchestrator only ever sees the returns. compaction never gets a chance to bite

the voice is always the hardest part to nail but the subagent orchestration here actually looks lethal

i just let mine spawn whatever it wants turns into a mess pretty quick

subagent controls are the real bottleneck now, finally glad to see them getting proper docs

Short format still gets those new subagent controls across clearly

the new features sound very useful

subagent controls in the orchestrator is the right layer. less babysitting mid-run.

hi teknium , i like use hermes to do all thing , but not sure why if code hard thing or long context it make ugly result. i try same model with codex it code better , did i do some thing wrong

what happens when subagents diverge mid-task and nobody notices

我的 Hermes 被公司的风控控制住了，说是有注入性风险。你们的项目代码是不是得在安全上好好做些审核？

The subagent controls look like a massive upgrade for the workflow. Don't worry about the voice, the functionality is what actually matters here.

steering subagents live is neat

The subagent controls look pretty useful

v0.21.0 shipped live mid-flight steering for subagents — not just spawn-and-wait. The orchestrator can redirect a running child without restarting the tree. Current tagged install is the v0.21.1 rollup (v2026.9.7); curated notes for that window land with v0.22.0.

Subagent control is a huge unlock

Subagent controls are useful if the defaults stay simple.

Is there a way u can get the Bots section to act more like a normal session? Specifically when i give Bots a task the main agent goes silent and seems to nevert show its thoughts or provide feedback it just goes silent. I never see my main bot/agent ever talk or call other bots

✓ Update complete! (v0.21.1) [main @ 1675f1f2c2]

on the serving side we learned every long-running agent turn needs a hard per-turn budget plus a watchdog that keeps re-arming, or zombie turns burn slots for up to 30 min. do the new subagent controls expose per-agent budgets (tokens/time), or is it lifecycle only?

Hi @Teknium, feature request: a tool that allows Hermes to not respond sometimes, whenever it isn’t necessary (if the user explicitly wants this). OpenClaw has this feature. Basically the ability to ignore a message that doesn’t require a response.

When subagents can fan out, send/money gates have to travel with them — otherwise blast radius scales with the orchestrator. Same war as skills fitness when the agent can touch customers.

"UH OH, IT STARTED CHARTING LIBRARIES" I like the voice, it's like a monotone grade 9 math teacher who really cares about her students. The kind of teacher who recognizes that you're a smart capable student, if only they could reach you and help you know your self worth.

Subagent controls in the orchestrator complete the picture. With them, your fleet stays coordinated, makes deliberate decisions, and scales with confidence.

Live steering is useful if intervention stays the exception, not the workflow

Sometimes I wonder if you guys are peeking into my secret local repos. First you broke my Hermes desktop app plugin based on the star graph, and now you are building in features and functionality I was working on weeks ago. Great minds think alike I suppose. Oh well. It's been a great learning experience for me, and that's where the value lies.

19 hours is wild man. focus on the subagent flow vs the context shredding debate is getting spicy today.

Isnt steer the default if you just type?

The interesting bit is treating subagents as steerable collaborators instead of fire-and-forget workers. Curious how Hermes decides when to interrupt vs let the helper finish—too much steering and the orchestrator becomes the bottleneck.

Can the orchestrator kill a runaway subagent directly?

Now this is what turns agent swarm into a super workflow