Mike Leiter

Systems Engineering. AI. Automation.
Are humans the meta-harness?

[Mike Leiter @MikeLeiterDev]

The emergence of Grok bot/Hermes bot swarm has really lit a fire under me!

I dove headfirst into agents last winter due to inspiration from OpenClaw, and have been in the middle of the agentic deluge ever since. New idea, frameworks, tools never stop popping up..

But a two working bot swarms have crystallized before our eyes in real time!

I see Grok/Hermes bot swarms as essentially the same as a Paperclip company. But these bots should not be reasoning over any hard requirements or workflows that can be put into deterministic code. Hard workflows need a hardcoded implementation - e.g. make a plan, make tests, implement, validate, review...

The top level bot swarm should just be deciding when to route a task into a hardcoded software or other workflow factory. Which doesn't mean there is no inference involved - a code factory has a coding agent at it's core. And maybe a planner bot, verification bot, etc. But the role of a workflow agent that is part of a known, well-defined process is fundamentally different from the role of a strategic-level, function-purposed bot like a technical director, software manager, or marketing manager bot.

The pieces are on the board. The required strategy has revealed itself. Now just to play the hand we've been dealt.

[Mike Leiter @MikeLeiterDev]

@bert_programmer Yes, the human governance layer is often not discussed.

For the moment, most can barely get past good Lv 3 agentic automation. Lv 4 is still early stages and brittle.

I'm not aware of any solid Lv 5 dark factories that are stable long-term, but maybe they aren't advertising..

[Mike Leiter @MikeLeiterDev]

The emergence of Grok bot/Hermes bot swarm has really lit a fire under me!

I dove headfirst into agents last winter due to inspiration from OpenClaw, and have been in the middle of the agentic deluge ever since. New idea, frameworks, tools never stop popping up..

But a two working bot swarms have crystallized before our eyes in real time!

I see Grok/Hermes bot swarms as essentially the same as a Paperclip company. But these bots should not be reasoning over any hard requirements or workflows that can be put into deterministic code. Hard workflows need a hardcoded implementation - e.g. make a plan, make tests, implement, validate, review...

The top level bot swarm should just be deciding when to route a task into a hardcoded software or other workflow factory. Which doesn't mean there is no inference involved - a code factory has a coding agent at it's core. And maybe a planner bot, verification bot, etc. But the role of a workflow agent that is part of a known, well-defined process is fundamentally different from the role of a strategic-level, function-purposed bot like a technical director, software manager, or marketing manager bot.

The pieces are on the board. The required strategy has revealed itself. Now just to play the hand we've been dealt.

[Mike Leiter @MikeLeiterDev]

@MichaelGannotti Consider running Hermes on the web from your persistent desktop. Here is my Android plan.

I already have Hermes connecting through Telegram on my phone, but after looking into desktop and the new bot capabilities I'm going to setup the web interface and see how that works.

[Mike Leiter @MikeLeiterDev]

I think the % of discussions about agent harnesses will drop because the average user will "just use AI" or will "just talk to their bot".

I don't think the amount of discussions about harness engineering will drop in absolute terms. I count it as the 3rd core AI agentic skill:

1. Prompt engineering
2. Context engineering
3. Harness engineering
4. Intent engineering

Power users will understand all 4, even if they are "under the hood".

[Mike Leiter @MikeLeiterDev]

@witcheer Symlinks (and sometime hard links/joins/aliases) are many times an elegant solution to an architectural problem.

Consider a similar approach for centralized skills that needs to be available to multiple different types of agent harnesses with different workspaces.

[Mike Leiter @MikeLeiterDev]

@BkashJosi Grok bot is the old Apple experience - simpler, few choices, but just works.

Hermes is Linux. You can customize everything. But you have to customize everything...

[Mike Leiter @MikeLeiterDev]

Seriously think about buying a used PC. Doesn't need much to be useful as a remote bot PC. If you want some local inference that doesn't need to be fast, consider an AMD v620 card or Tesla v100. Both 1/3 to 1/2 of a 3090, but 32 GB cards.

They fit really nice into a used corporate PC like a P520 or Z4 G4, if you know what to look for. If you have a spare SSD and spare video card you can probably get one and 32 GB memory for under $300. If you need those, under $400. Linux will run really nice on it...

https://t.co/iB93a2kb6N

[Mike Leiter @MikeLeiterDev]

I just received my $200 used ThinkStation P520. The ECC memory is still coming, as is the Tesla V100 32GB inference card and a fan assembly to give it cooling.

And the PSU and MB can support a 2nd full length card up to 250 W... but I'll test the first card first.

All together - just under $1000 for a full inference PC that runs about half the speed of a 3090, but with more memory.

I'll report how things go with Qwen3.8-27B, among other planned tests.

[Mike Leiter @MikeLeiterDev]

@unclebobmartin Standard followup to anything my agent does:

"Critically evaluate <x>, surfacing gaps and concerns".

Doesn't matter if it's the same model/agent that just generated <x> or not!

[Mike Leiter @MikeLeiterDev]

I have yet to find a real use for it, though the 20% price reduction gave a bit more separation from Sol, which was needed, but it's not really far enough, given how far Luna was reduced.

I'd rather use Grok 4.6 or even 4.5, or now maybe Gemini 3.7, for anything that Luna can't handle but I don't need Sol or Opus 5 for.

[Mike Leiter @MikeLeiterDev]

So, traditionally, this means worktrees for each agent - they have their own branch an a worktree just ends up being a separate folder with a branch working directory in it. You only need one main repo store - it can be a traditional clone or a bare-repo (object store only), with the right references.

This will allow them to work independently in parallel, just as though they have independent clones of the repo, but only needing the object store once.

Any conflicts will need to be resolved at merge time. Of course, you will still have conflicts and need to resolve them - but you can have them cheat a bit if you want by having them peek at the other worktrees, or ensuring the architectural planning is aware of all the ongoing work somehow. In part, this is what a good swarm setup with individual bot communication channels helps to address. Both Grok bot and the new Hermes bot swarm have agent communication channels you can setup and control.

[Mike Leiter @MikeLeiterDev]

@bios_hazard I can only make it stop by getting you to think about something even more ridiculous and annoying to have trapped in a loop in your head.

Or is it trapped in a repeating graph? 🤣

[Mike Leiter @MikeLeiterDev]

@HermesWatcher This is getting crazy! Time to upgrade Hermes and try all this out.

Any tips so far?

[Mike Leiter @MikeLeiterDev]

Wow, put this together with a list of all the features that Hermes has put together for it's increasingly-capable bots, and it looks like it's increasing what it can do at an increasing rate!

Time to upgrade to the latest and try this out...

[Hermes Release Watch @HermesWatcher]

Hermes had another ridiculous week.

Last week was huge. Somehow, it did not slow down.

In the last seven days, Hermes learned to control the browser you are actually looking at, Bot Mode turned into a multi-machine agent fleet, conversations gained live interactive UI, Cron got persistent memory, and `/review` gave finished work an independent second set of eyes.

And those are just the headliners.

Here are the biggest Hermes changes from August 17-23:

▸ Hermes can now use the browser you are actually looking at:

The in-app Preview browser is no longer something Hermes can only open and read.

Hermes can inspect the page, click, type, scroll, press keys, navigate, reload, and even annotate an element so you can see exactly what it is targeting.

That means Hermes can work inside the visible browser session you are already signed into while you watch it happen.

This opens up a ridiculous number of workflows.

▸ Bot Mode became a real multi-machine agent fleet:

A lot of what was coming together last week became real infrastructure this week.

Desktop can now give you one Bot Mode roster across registered Hermes gateways while keeping every bot tied to the machine that actually owns it.

Local. Remote. SSH. Hermes Cloud. Different machines.

Remote bots can open their real Bot Chat without moving your normal Sessions workspace.

Bots also gained `message_agent`, giving them an actual agent-native way to message teammates instead of constructing shell commands.

And bots across different Desktop connections can message each other too.

Group rooms also gained real threads plus images, PDFs, files, paste, and drag-and-drop attachments.

Bot Mode is starting to look a lot less like “multiple profiles” and a lot more like an actual team.

▸ Your Hermes conversation can now become an interface:

Plugins can render live UI directly inside the conversation.

Not a screenshot.

Not a link to another app.

Actual interactive UI inside the transcript.

And it does not have to be one-way.

Interacting with that UI can send a real turn back to Hermes behind the scenes, let the agent do more work, and update what you see.

Chat does not have to end in text anymore.

▸ `/review` gives your agent an independent second set of eyes:

Your main agent finishes the work.

Run `/review` and Hermes launches a separate background reviewer with its own tools to inspect the actual work behind the conversation.

Code. PRs. Docs. Research. Other referenced artifacts.

The reviewer can work under the repo’s own project instructions and can be told which skills the primary agent was using.

You can also give Review its own model.

One model does the work.

Another independently checks it.

The findings return to the original session so the main agent can respond, defend the work, or fix what it missed.

This is a big one.

▸ Cron became much more than “run this prompt at 8 AM”:

Cron agents can now use Hermes persistent memory, including MEMORY.md and USER.md.

Individual jobs can also pin their own reasoning effort, so a lightweight recurring check does not need the same reasoning budget as a scheduled deep analysis.

And scheduled output can now be delivered directly into a bot’s canonical Bot Chat as a real incoming turn.

The bot reads it and responds.

Now start combining scheduling, memory, Bot Mode, and unattended work.

That gets interesting very quickly.

▸ `hermes update` now checks whether the update actually worked:

This one is less flashy, but it matters.

Hermes has been rebuilding the updater around the reality that one installation may have multiple profiles, gateways, services, and runtimes running at once.

`hermes update --plan` can inventory that fleet before touching anything.

Updates generate structured receipts.

Running gateways can report which code they are actually serving.

And the restart phase now has to account for the runtimes that were in the plan.

If Hermes expected something to be updated and it was missed, stale, or down, the updater can fail visibly instead of quietly declaring victory.

That is the kind of boring infrastructure improvement you really appreciate the first time something goes wrong.

▸ Long-running jobs got a serious reliability pass:

The old default 500-turn ceiling is gone.

Hermes now runs with unlimited turns by default unless you choose to set a cap.

New stall guards also look for agents repeatedly making the same tool call or announcing that they are about to continue and then stopping.

And if Hermes runs the same tool again and gets the exact same giant result, it still executes the tool fresh, but it can feed the model a tiny reference to the previous result instead of dumping another 20K-50K characters into context.

MCP results also got tighter context handling.

The goal is pretty simple:

Let long jobs keep working without letting the agent quietly spin its wheels or bury itself in repeated context.

▸ Fresh Hermes installs need fewer API keys before they can do useful work:

Hermes now has keyless paths for web tooling across multiple supported providers.

There is also a new `opencode-free` provider for OpenCode’s free tier that does not require an API key or OpenCode account.

It appears in the normal model pickers too.

Install Hermes.

Pick an available free model.

Give it a job.

The distance between “I just installed Hermes” and “my agent is doing useful work” keeps getting shorter.

▸ Repositories can now bring their own Hermes skills:

A repo can include project-specific skills under `.hermes/skills/` or `.agents/skills/`.

Inside that project, those skills can take priority over lower-level versions of the same skill.

Hermes also puts a trust boundary around this.

The repo has to be trusted, project skills are scanned when loaded, and an unknown repo does not silently get permission to inject its own agent instructions.

For teams building repeatable Hermes workflows around a codebase, this could become extremely useful.

Also worth knowing:

→ Desktop voice can use supported STT/TTS providers directly through the active profile.

→ Eligible Codex GPT sessions now default to 272K context, with explicit `-900k` variants when you actually want the larger window.

→ `/model` got fuzzy search.

→ Desktop got a command palette.

→ `/status` got much more useful session information.

→ `clarify` can ask several independent questions in one batch instead of interrupting you one question at a time.

→ Hermes can apply Desktop layout presets itself.

→ Plugins can be installed from Git or `hermes://` links with a review step before anything gets installed.

Last week’s story was Hermes learning to keep working, delegate work, supervise agents, and coordinate across machines.

This week is what happens when those pieces start connecting.

The agent can use the browser you see.

Bots can work and communicate across machines.

Scheduled jobs can remember, wake bots, and hand work back into the team.

Agents can build interfaces inside the conversation.

And when the work is finished, another agent can independently inspect it.

Hermes is starting to look less like one very capable agent with a long feature list and more like an operating layer for a team of agents that can actually keep working together.

[Mike Leiter @MikeLeiterDev]

"...critically think along with it..."

There's the key. Critical thinking.

Search for what is good, true, correct, and real. Understand the AI is just estimating words synthesized from what it's been trained on plus what the harness puts in it's context. You must press it and probe it's answers for how it concluded that, how it resolved any conflicts or priorities, what it's sources and confidence is - all the things you also do with humans presenting you conclusions.

Those who get wow'd by it's breadth of knowledge are lost.

[Mike Leiter @MikeLeiterDev]

@tonysimons_ When can I run Hermes Desktop on my Android?

I will pay double for this! 😉😇

[Mike Leiter @MikeLeiterDev]

If you haven't ever created an AGENTS.md file to guide your agent, or not sure you've got it optimized, I recommend checking this out.

[Kun Chen @kunchenguid]

many people asked me how to write CLAUDE.md or AGENTS.md, and i see lots of bad advice flying around

so i took some time to write down a guide in https://t.co/v9rrkWKEFr

tl;dr
- handwrite your user level AGENTS.md
- for project level ones, you don't write it. you train it like a neural net

i also open sourced my private solution "backpass" at https://t.co/DkM9b4TcZ0 - it samples your past agent sessions for a repo, distill key learnings and losses, synthesize them, and produce a gradient descent step as a proposal that you can review and apply to improve your AGENTS.md and project level skills

easiest way to run it is just "npx -y backpass" in your repo

hope it helps! please share with whoever you think can benefit from it

[Mike Leiter @MikeLeiterDev]

Ha! You are quite a ways ahead of me here. I'm just far enough in to be dangerous.

But with the v100 coming in I decided I'm going to need a proper benchmark suite and maybe going through your feed and having AI explain it to me will get me better up to speed. Current plan is Phoenix as the core of that, probably with Inspect on the top (just installed Phoenix this week and haven't done anything but collect traces so far). Have any good benchmarking tools I should look at, especially for local models?

Every time I ask AI a proper Socratic question about local models and benchmarking, I find out a new aspect it's been holding back.. though to be fair, there's new things happening every week.

[Mike Leiter @MikeLeiterDev]

@SaintsOfCompute @venturetwins That's fair. I guess I believe in being a life-long learner. I'm not always the best at it, but I do my best. AI is the best opportunity I've seen to x10 my skill.