Hermes Agent Goals (NEW Persistent Update FREE)

The Hermes agent goals update is the biggest free upgrade to any open-source AI agent in 2026, and after running it for the past few weeks I'm convinced this is the moment when truly autonomous multi-step work becomes practical for solo founders. One command kicks off a goal, Hermes chases it across multiple turns, and a judge model verifies completion — no babysitting required.

This post breaks down exactly what Hermes goals does, how to install and use it, the four commands you need to know, and the use cases where this feature genuinely earns its place in your stack.

🔥 Want my Hermes goals 30-day roadmap + use cases? AI Profit Boardroom has the full goals roadmap + weekly live coaching. → Get the roadmap

What Hermes Agent Goals Does

You give Hermes one command and it chases the goal step by step until completion. The pattern is inspired by the Ralph Loops approach popular in the Claude Code community, where a judge model checks after each turn whether the goal is done. If not, the agent continues working. The loop runs until the goal is achieved, the turn budget is exhausted, or you manually pause or clear it.

What this means in practice is that you can stop micro-prompting agents through every step and instead set a high-level outcome and let the agent figure out the path.

How To Install

Open your terminal and run:

hermes update

That pulls the latest features including persistent goals. The update takes seconds and doesn't require any reconfiguration.

The Four Commands

Four commands cover the entire goals workflow.

The /goal <text> command sets a persistent goal — for example, /goal build out an SEO website with 10 articles. The /goal status command shows the current state of the active goal so you can see what Hermes is working on. The /goal pause and /goal resume commands let you pause work, close your terminal, and pick up tomorrow without losing context. The /goal clear command ends the current goal entirely if you want to abort or start fresh.

Together these four commands give you full control over the autonomous loop.

Watch The Walkthrough

For Hermes basics if you're new to the framework, the walkthrough below covers the foundations.

How The Loop Works

The loop has three layers that work together to keep autonomous work coherent.

The first layer is goal acceptance. When you set a goal, Hermes confirms what it's been asked to do and initialises a planning step so it has a shape for the work ahead. The second layer is the turn loop. Each turn is a normal Hermes action — code, file edit, web search, or whatever the next step requires. The third layer is the judge check. After each turn, a judge model evaluates whether the goal is done. If yes, the loop ends with a goal-achieved message. If no, the agent continues into the next turn.

A turn budget (default 20 turns, configurable) prevents runaway loops if something goes wrong with the judge or the work.

Configurable Settings

Two main settings to know.

The goal_max_turns setting controls the maximum turns before the loop stops. Use a bigger budget for bigger goals — a 50-step task obviously needs more than the default 20 turns. The goal_judge setting controls which provider and model is used to judge completion. Use a stronger judge model for nuanced goals where the criteria are harder to evaluate.

Best Use Cases For Hermes Agent Goals

Six use cases where this feature delivers genuine leverage.

The first is building full websites with a single command like /goal build out a 10-page SEO website on [topic] and deploy to Netlify. The second is multi-step research with prompts like /goal research [topic] across 10 sources + write a comprehensive summary. The third is code refactors with goals like /goal refactor this codebase to use TypeScript + add tests. The fourth is content series — /goal create a 7-day email series on [topic] will produce the entire sequence end-to-end. The fifth is data analysis where you say /goal analyse this dataset + produce a structured report and Hermes handles the whole thing. The sixth is bug squashing with /goal find and fix every bug in this repo.

For each of these, one command kicks off autonomous multi-step work that used to require hours of manual prompting.

What Makes A Good Goal Prompt

Three rules for writing goals that actually complete reliably.

The first rule is specificity. "Build a website" is too vague for the judge to evaluate, while "build a 10-page SEO website on the Hermes agent topic, dark theme, deploy to Netlify" gives both the agent and the judge clear targets. The second rule is measurability — the judge needs concrete completion criteria like "with at least 10 published articles plus sitemap.xml live." The third rule is bounded scope. Keep the goal within reasonable scope for your turn budget; if a task genuinely needs 100 turns, set the budget accordingly upfront.

Resume Across Sessions

This is one of the most powerful features of the goals system.

Set a goal on Monday, close your laptop on Tuesday, and run /goal resume on Wednesday. Hermes continues exactly where it left off, with full context preserved. For multi-day projects this is the unlock that turns the agent into something genuinely autonomous rather than tied to a single working session.

Common Goals Mistakes

Three mistakes I see people make when they first try goals.

The first is vague goals where the judge can't determine completion. Be specific about what "done" means. The second is using the default turn budget for big goals — 20 turns won't finish a 50-step task, so adjust the config when scope demands it. The third is ignoring the goal status. Check status periodically, and if the agent is looping unnecessarily, clear and re-prompt with sharper specifics.

Pairing With Other Hermes Features

Three combinations that compound when paired with goals.

Goals plus skills works well when the goal triggers and orchestrates specific skills that handle individual steps. Goals plus memory persistence (see OpenClaw Memory Persistence for related concepts) keeps long-running goals coherent across days. Goals plus multi-agent setups let a manager agent set sub-goals for worker agents, which is the basis for the swarm patterns I cover in Hermes Agent Swarm.

Cost Of Running Goals

The economics depend on whether you're running local or cloud.

Local Hermes via Ollama is free regardless of how long the goal runs. Cloud LLM-backed work has per-token costs that scale with goal length. For a typical goal you'll spend £0.50-£5. For multi-day goals expect £5-£20. That's cheap relative to the value when the alternative is hours of your time on micro-prompting.

Real Goals I've Run

Five real goals from my own work to give you a sense of what works.

Goal one was building this site — a 25-article SEO site that took 20 turns to first version and 40 turns to polished and deployed. Goal two is a daily research roundup of the week's AI news, which takes about 5 turns each morning. Goal three was a code refactor of a bigger codebase that ran 50+ turns over two days. Goal four was a 7-day email content series that completed in 20 turns. Goal five was a quarterly analytics deep-dive that took 30 turns.

All five ran via a single /goal command with no babysitting between turns.

🚀 Want hands-on goals coaching? AI Profit Boardroom has weekly live coaching to wire goals into your workflow. → Join here

Goals Vs Manual Prompting

Aspect	Manual prompting	Hermes goals
Engagement	High (every step)	Low (one command)
Coherence	Drift risk	Judge keeps on track
Time	High	Low
Reliability	Depends on you	Depends on judge model
Multi-day	Hard	Resume command

For multi-step work, goals win on every dimension that matters except control.

Compounding Effect

Daily use of goals saves hours per day, and the savings compound over time.

In week one you'll save roughly five hours. By month one you've saved 20-30 hours. By the end of quarter one you've saved 60-90 hours. That's a meaningful productivity unlock, especially for solo founders where every recovered hour can go straight into higher-leverage work.

Common Member Questions

Three questions that come up repeatedly from members.

"Is the judge model accurate?" Generally yes, and for nuanced goals you can swap in a stronger judge model via config. "Can goals run overnight?" Yes — they'll run until the budget is exhausted or the goal is completed, so you can kick one off before bed. "What if the goal stalls?" The goal status command will show what's happening, and you can clear and re-prompt with sharper specifics.

Installing Today

Three steps to get up and running today.

Step one is running hermes update to pull the goals feature. Step two is trying /goal with a small task like "build me a markdown file with five ideas on X" so you get a feel for the loop. Step three is iterating — try bigger goals, adjust the turn budget config, and build the muscle for writing good goal prompts.

When To Use Goals Vs Skills

Both have their place in a well-built stack.

Use goals for open-ended multi-step work where the path isn't fully known upfront. Use skills for repeated patterns that look the same every time you run them. The combination is powerful — a goal can trigger skills as part of its execution, giving you both the autonomy of goals and the reliability of well-defined skills.

Quality Of Output

Honest assessment of what goals does well and where it struggles.

The strengths are persistence across long tasks, multi-step coherence that holds up better than manual prompting, and the judge model keeping the agent on track when it would otherwise drift. The weaknesses are that the agent can drift on ambiguous goals where the judge isn't sure what "done" looks like, and that big goals need careful framing to complete reliably.

For most professional use cases, the output quality is excellent.

FAQ — Hermes Agent Goals

Is it really free?

Yes — Hermes is open source and the goals update is part of the free release.

How do I update?

Run hermes update in your terminal. Takes seconds.

What's a good first goal?

Start with a small task that completes in 5-10 turns so you can see the loop work end-to-end before tackling bigger goals.

Can goals fail?

Yes — most commonly when the turn budget is exhausted before the agent finishes. Increase the budget or sharpen the goal.

What does it cost to run locally?

Free with Ollama. Cloud-backed runs have per-token costs.

Best paired feature?

Memory plus multi-agent for the most powerful setups. Skills for reliability on common patterns.

Worth it for solo work?

Yes — it eliminates the micro-prompting that takes most of your day.