Hermes Delegate Task: How v0.21.0 Turns One Agent Into a Team

The hermes delegate task capability — the delegate_task tool inside Hermes Agent — is how one Hermes agent hands work to child agents and runs them like a small team: you give the parent a goal, it spins up subagents, and as of Hermes Agent v0.21.0 (released 31 August 2026 by Nous Research) you can watch those children live, steer them mid-flight with course corrections, and stop them early while keeping the partial results they have already produced. What you get, according to the official v0.21.0 release notes on GitHub, is genuine live orchestration rather than fire-and-forget delegation — plus JSON-schema validation on child outputs, per-delegation cost reporting, and raised defaults of 250 iterations and 10 concurrent children. For anyone automating real business work with agents, this is the single most practical upgrade in the release.

📺 Watch: Give Hermes One Idea… It Does the Rest

🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside · Want AI SEO help 1-on-1? Book a free SEO strategy session →

This guide walks through what delegate_task does, what changed in v0.21.0, how to structure delegations so the output is actually usable, and where delegation fits alongside swarms and scheduled jobs in a working automation stack. The wider release is covered in the Hermes v0.21 overview; this page goes deep on the delegation piece only.

What Delegating a Task in Hermes Actually Means

When a Hermes agent calls delegate_task, it creates a child agent with its own instructions, its own context and its own iteration budget, pointed at one specific job. The parent carries on with the main thread while children work in parallel. That structure matters for automation because it mirrors how you would brief a human team: one coordinator, several specialists, each with a clear deliverable. Delegation is what stops a single agent context from drowning in research notes, drafts and tool output — the child does the messy work, the parent receives the finished result.

Before v0.21.0, the pattern worked but ran mostly blind: you briefed a child and waited to see what came back. The v0.21.0 release notes describe the change as live orchestration, and the difference shows up in three specific abilities: you can list running children, steer them mid-flight, and stop them early while preserving partial results.

The Hermes Delegate Task Upgrade in v0.21.0, Point by Point

• List running children. The parent can enumerate what is currently delegated and running — essential once several children work concurrently and you need to know what is actually in flight.
• Steer mid-flight. Course corrections reach a child while it works. Spot a child researching the wrong competitor two minutes in, and you redirect it there and then instead of discovering the waste twenty minutes later.
• Stop early, keep the partials. Ending a child no longer throws away its work; partial results are preserved. A research child that has covered eight of ten sources still hands you those eight.
• JSON-schema validation on outputs. Optionally declare the shape a child must return. If your pipeline expects a list of headlines with URLs, validation catches a rambling essay before it breaks the next step.
• Per-delegation cost surfacing. Each delegation reports what it cost, so you can see which children earn their keep — the missing piece for anyone running delegation at volume.
• Raised defaults. Limits now default to 250 iterations and 10 concurrent children, per the release notes, which is enough headroom for serious multi-step pipelines without touching configuration.

If you want delegation pipelines that already work — research teams, content teams, SEO agents briefed and scoped for you — the AI Profit Boardroom includes the full Agent OS plus five live coaching calls a week to get yours running. Rather talk strategy first? Book a free SEO strategy session.

How to Structure a Delegation That Comes Back Useful

1. One deliverable per child. "Research this topic" produces mush; "return the five most-cited sources on X with one-line summaries" produces something the parent can use. The delegation is only as good as the brief.
2. Declare the output shape. With schema validation now available, define the JSON you want back for anything feeding another automation step. Free-form text is for humans; structured output is for pipelines.
3. Set children up to fail fast. Give each child the narrowest context that lets it succeed. Smaller briefs mean earlier, cheaper failures — and mid-flight steering means you can correct instead of restart.
4. Watch the costs per child. Use the new cost surfacing to find the expensive delegations. A child that burns most of your spend to produce a summary you could template is a child you should replace with a cheaper model or a fixed workflow.
5. Keep the parent thin. The coordinator's job is briefing, steering and assembling. If the parent starts doing the specialist work itself, you have lost the benefit of the structure.

📺 Watch: New Hermes Agent Update is SCARY GOOD!

Delegation vs Swarms vs Cron: Picking the Right Tool

Hermes gives you three ways to multiply one agent into many, and mixing them up is the most common automation mistake. Delegation via delegate_task is hierarchical and session-scoped: a parent, its children, one goal, then everyone goes home. A swarm — covered properly in the Hermes agent swarm guide and the broader Hermes swarm write-up — is a set of longer-lived agents working loosely in parallel on an ongoing area. Scheduled jobs through the Hermes cron system handle recurring work on a timer, and v0.21.0 separately gave those jobs persistent memory between runs. The rule of thumb: cron for "every day", swarm for "always on", delegate_task for "big job, right now, needs coordination".

Delegation also composes with the rest of the stack. Children can use the same tool servers as the parent — see the Hermes MCP server guide for how tools are wired in — and the patterns in the Hermes agent use cases round-up show where multi-agent structures actually pay off in practice.

Why This Matters for Your Automation Stack

For Hermes users, the gap between a demo agent and a production agent has always been control: what happens when the work goes sideways at step forty. With v0.21.0, delegation gains the three controls production work demands — visibility (list children), correction (steer mid-flight) and a safe exit (stop early, keep partials) — plus the validation and cost data to run it accountably. That is the boring, load-bearing infrastructure that lets you hand an agent team a real deliverable with a deadline.

If you are building towards that, two resources on this site pair well with this guide: the Agent OS write-up covers the instruction-file layer that makes parent and child agents behave consistently, and the Goldie Bench write-up covers how the candidate models compare in hands-on tests — useful when deciding which model your children should run, since the cheapest adequate model per child is where delegation costs are won.

Hermes Delegate Task: Quick Answers

How many subagents can Hermes run at once?

The v0.21.0 defaults are 10 concurrent children and a 250-iteration budget, per the official release notes — raised specifically so that serious pipelines work without configuration changes. Treat those as ceilings for one delegation wave, not targets: most well-designed automations use three to five children with tight briefs and finish faster than a ten-child free-for-all.

Can you stop a delegated task without losing its work?

Yes. Stopping a child early now preserves its partial results, which changes how aggressively you can prune. If a delegation is drifting, you stop it, keep the useful fraction, and re-brief — rather than choosing between letting it burn budget or losing everything it produced.

What does delegation cost?

Whatever the child agents consume in model usage — and v0.21.0 finally makes that visible with per-delegation cost surfacing in the results. There is no separate fee for the mechanism itself; the practical cost control is putting cheaper models behind high-volume children and reserving your strongest model for the parent.

How is delegate_task different from Bot Mode?

Same release, different shape of teamwork. Bot Mode, also new in v0.21.0, gives you named agents with faces and group chats in the desktop app — persistent characters you talk to. The delegate_task tool is programmatic and session-scoped: a parent creates workers for one job and dissolves them when it ships. Use Bot Mode for an ongoing team feel, delegation for coordinated bursts of real work.

If you want a working agent team without months of trial and error, check out the AI Profit Boardroom — 1,000+ done-for-you workflows, the Agent OS and live coaching every week. Prefer a personal roadmap? Book a free SEO strategy session and map it in one call.

Real wins from inside the AI Profit Boardroom

See all 3,000+ members →

← Back to all posts