Published: 2026-05-05

Hermes Agent Curator: Automatic Agent Selection and Task Chaining

Chapters / key moments (click to jump — plays here on the page)

Hermes Agent v1.3 ships the Curator feature — a system that takes a task description, breaks it into subtasks, scores every available agent and tool against each subtask, assembles the best team, and runs the whole pipeline in sequence. You describe what you want; Curator decides which agents are best suited, which model each should run, and how to chain their outputs. The agent picker learns from your feedback over time, improving its selections the more you use it.

Source video

"Hermes Agent Curator is INSANE!" by Julian Goldie SEO — Watch on YouTube →

Key Takeaways

Curator removes the need to manually select agents, tools, and models for each task — you describe what you want, and the system builds the team and runs the job.
Three internal components: a task reader (understands the full intent, not just the surface ask), an agent picker (scores agents, tools, and models against each subtask), and a runner (chains outputs from one agent to the next).
The agent picker learns from your acceptance and rejection patterns — the more you use Curator, the better its selections get for your specific workflow.
Reported speed improvement: tasks that previously took 30 minutes now take around 5, primarily because the right agent is selected on the first try rather than after failed attempts.
Works with any model Hermes supports: Claude, GPT, Gemini, and local models via Ollama.
Hermes Agent is open-source and runs locally — Curator runs entirely on your own infrastructure.

How Curator Works: Three Components

The task reader parses your prompt to understand not just the words but the underlying goal. It distinguishes between a surface request ("write social posts") and the full job ("research what's trending in the space, draft posts for the next 30 days, schedule them across platforms"). The agent picker then looks at every agent you have configured, every tool available, and every model option, and scores them for each subtask. It picks the fastest and cheapest appropriate option, not just the most capable one. The runner executes the agents in order and passes each agent's output as input to the next — a research agent's findings go directly to a writer agent's context, which go directly to a reviewer agent, all without your involvement.

Practical Example: Monthly Content Plan

The demo use case: tell Curator to build a full content plan for next month. Curator deploys a research agent to find what's trending in your space, a writer agent to draft the posts, a planner agent to map publication dates, and a reviewer agent to check the work — all triggered by a single task prompt. This kind of multi-agent pipeline previously required manually configuring each agent, knowing which agent type handled which job, and sequencing the runs yourself. Curator collapses that setup into one step.

Related on OpenClawDatabase

Hermes Agent Setup Guide — getting Hermes running on your VPS or Mac
Hermes Tasks — scheduling and running automated tasks
Hermes Memory — how Hermes stores and retrieves context
Compare Agents — OpenClaw vs Hermes and other platforms

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

What curator replaces
The old flow was: pick your agent, pick the tools, pick the model, write the prompt, set the task, go — and if you picked the wrong agent, you get a bad output and start over. <strong>Curator takes the goal and makes those selections itself.</strong>
Three parts, in order
<strong>The task reader</strong> parses what you wrote and tries to identify the real goal rather than the surface ask. <strong>The agent picker</strong> scores every agent, tool and model you have configured — which agent suits this job, which tool helps most, which model is fastest and cheapest for this kind of work. <strong>The runner</strong> executes them in order, feeding each output into the next input.
It learns from your corrections
The picker observes which agents you keep and which you reject, which jobs went well and which failed, and adjusts. There is a <strong>feedback button built into the tool</strong> — thumbs down on a bad pick and the next selection improves. You are training a router over time.
It layers onto the existing task board
If you already run the multi-agent Kanban board, curator slots on top as <strong>the brain that decides which agent works which card</strong> — you stop assigning agents to cards manually. A single high-level card gets broken into smaller cards, each assigned and run.

Gotchas

<strong>This is a promotional video for the presenter's paid community, delivered by an AI avatar, with the pitch in the middle.</strong> The feature description is checkable against the release; the framing is marketing.
<strong>The honest limitation is stated in the video and is the most useful line in it:</strong> "curator is only as good as the agents you give it. If you only have one agent, it picks that one every time." The setup work does not disappear — it stops being repeated per task.
Also stated: it does not always pick correctly, and the team behind it is open about that, which is why the feedback control exists.
The "30 minutes becomes five" figure is attributed to unnamed testers and is not measured here.

← Back to News digest · See also: Hermes guide