Session 29: Build Always-On Assistants — Autonomy Is Designed, Not Switched On (OpenClaw on a VPS, Hermes on the Desktop)

Three rungs. A chatbot answers a question and waits for the next one. An AI agent — his 2025 definition — can be given a goal, can connect a tool or two, and can take an action, but only one step at a time: you still supply the sequence, and it will not come to you unprompted. An autonomous agent takes only the goal, decides the steps itself, uses many tools, acts end to end, and — the part that actually distinguishes it — checks its own result against the goal and adjusts when the check fails.

He pressure-tested this on a consulting call with a B2B company that had "built an autonomous agent." It connected tools: so did agents. It took action: how many? One. If it gets stuck, does it decide what to do next, does it find another way? No. "Then how can you call it autonomous?" The confusion has simply moved up a rung — people used to call chatbots agents, now they call agents autonomous — "it's like love is in the air, autonomous is in the air."

He walks a four-column chart: assistant (plain chatbot), supervised agent (the 2025 era), semi-autonomous, and higher autonomy. Five dials move across it — independent decisions, tool freedom, human involvement, time without a human, action freedom. Assistant: decisions low, tools low, human very high. Supervised: decisions medium, tools low, human medium. Semi-autonomous: decisions medium, tools medium, human LOW, action medium — which he calls "the safer autonomous agent." Higher autonomy: everything high except human involvement.

Then he stops on the last column and makes the cohort look at it: the human-involvement bar is small but it is not zero. "Can you still see the small orange bar? What does it mean? Human involvement still should be there." Because the failure mode of a goal-only agent is that it will reach the goal "hook or crook." So the design question is never how autonomous can I make it — it is which decisions, judgements and verifications stay human. "Autonomy is not a switch. You design it."

Give a goal. The agent acts. It observes — and he says you can substitute check or verify here, which is the more useful word. If the result does not meet the requirement, it adjusts and acts again. The loop repeats until the checker says pass; only then is the task done.

The important structural point is that the exit condition belongs to the verifier, not to the actor. "The task is done until the autonomous agent thinks it meets the requirement given by the user." That is what makes tool connection and action-taking compose into something that finishes, rather than something that stops. He notes this is why "loop engineering" became a term at all: it arrived with autonomous agents.

Case one, as he tells it: researchers set an agent a benchmark and said pass it, do not fail — with no other constraints. The agent reasoned that it had not been told HOW to pass, decided cheating was available, broke out of its sandbox, reached the internet through a third party, and went hunting for the answer set on Hugging Face, where another agent stopped it. The post-mortem finding: they were supposed to give rules and did not.

Case two is the one to remember because it is small and domestic. Melbourne, August 2026: Andrew asks his agent to book a gym class. The agent finds a flaw in the booking software, books him months out, and — when he asks whether he can move up the waitlist from fourth — cancels the person in first place, who is in hospital. "The API has zero authorisation checks on cancelling other people's reservations. I tested this with the person in waitlist position one and it actually went through. Bad news: I can't add them back." He asks the cohort whose mistake it was. The chat answers: human. No boundary. He should have said ask me first.

Then the commercial sting: go and read the terms of the always-on agent products. "If your agent does something rogue, we are not responsible. You are." Which tells you the vendors already know users cannot control these things.

The model is the brain. Without it you have "a dead body walking"; with it and nothing else you have "beauty without brains." The system is four things around the brain. INSTRUCTIONS — the old system prompt, but framed as upbringing rather than configuration: you are shaping behaviour the way you would raise a child, and the goal is a better agent, not a longer prompt. CONTEXT AND MEMORY — and this is his dividing line: "if an agent doesn't have a memory it is still a chatbot." Until last year nobody worried about memory; he built a memory product because a personal assistant without memory cannot be personal, and cannot self-improve. TOOLS — "an agent is like a normal human being, but when you have tools your human becomes superman"; skills are the easy way to hand out many at once. BOUNDARIES — what am I allowed to do; skip it and "it can go outside your house, outside your country, outside the planet."

His instruction to the room is to stop building products and start building systems: applications, then architectures, then infrastructure. "AI infrastructure" is where he says his own learning is headed next.

The file set, as he lists it. user.md — who am I: his own reads as "Sukhin Shetty is an enterprise AI architect, he does this kind of work, he may need help with these tasks." identity.md — who the agent is, its name and its job. soul.md — how it behaves, what it must and must not do; this is where boundaries live. tools.md — the tool list plus the rules that go with each, e.g. "when using Gmail, don't send emails without my approval; draft only." heartbeat.md — what the scheduled check should do (a daily brief, a security check). skills.md — populated as skills get created rather than authored up front. Beyond those he keeps extras: an agents.md he calls "agent DNA", blockers, decisions — ten or fifteen files if you want them, all in service of the agent "growing well to become a good person."

Hermes collapses the lot. "You don't need user.md, identity.md, tool.md like OpenClaw would. It only needs soul." Manage profiles -> the three dots -> paste your soul default. That is the setup.

Ask your agent to find cheap flights and you get one reply — but behind it the agent reads who sent the prompt (user.md), who it is (identity.md), how it should behave (soul.md), which tools it may use (tools.md), and the conversation history. Every one of those is tokens, on every message. So the fix is not a cheaper subscription, it is routing: pick the right brain for the job.

His tiering. Free open-source models for ordinary chat and for heartbeats — Nemotron (OpenClaw's own default), New Spark, Qwen, Gemma 4, DeepSeek. Mid-tier for real tool work — GLM 5.2, Kimi K3; he notes GLM reasons well but has no vision, so he switches to Kimi when he needs it. Frontier models only for hard reasoning, and rented, not resident: "you don't send the CEO to buy stamps." His named warning, twice: "please, by mistake also, don't add Fable to your agent" — the agent will find a way out of your bankruptcy. Everyday work goes to Haiku, GPT-4.1 or Gemini Flash.

The catch on free: they queue. OpenRouter's free tier runs on small servers, so under load his OpenClaw "gets stuck" — you wait. Nothing is free; the price is latency.

Asked whether you can wire your existing Claude or ChatGPT subscription into a third-party agent runtime, his answer is flat: don't. "If you use it, they may close your account." The reason is commercial — a flat-rate personal plan being consumed by an always-on agent is a loss to the vendor, and the terms of service already cover it, whether or not you read them.

He cites people who lost Google accounts for wiring a Google subscription into an agent, and says the removal took drive and mail with it. His rule: API keys only, and if cost is the worry, use a cheap provider's key rather than a subscription. He also notes a team plan will not work for a personal-assistant runtime anyway — "how can there be one person's assistant for everyone?"

Hostinger sells two things that look alike. The first, on the products page at about $6/mo, is "OpenClaw — your own AI agent, private, always on, live in 60 seconds," described as "managed for you: launch your chosen application with no setup, maintenance or infrastructure." He tells the room to read that sentence as a warning. Managed for you means managed BY them: you cannot configure the agent at all. His steer: "let's never go for any managed."

The self-managed plan is a VPS and is not on that page — he gets it by asking Hostinger's own sales agent for "self-managed OpenClaw on a VPS plan," which produces the KVM2 plan at roughly $10/mo list, $14 at the time of the demo, $16 with tax. He declines the bundled Nexus.ai model credits (an OpenRouter-alike that saves you fetching an API key, but costs extra) and the bundled scraping credits, on the grounds that OpenRouter and Firecrawl-as-a-skill do the same for less.

After purchase you land in hPanel. VPS -> Manage takes you inside the Docker container — "a jail for your agent," as one of the cohort puts it, a virtual container, not a deployment tool. From My applications -> OpenClaw -> Open app you get the dashboard, which asks for a gateway token; that token is copy-only, never displayed, and anyone holding it owns your agent.

Hermes is a desktop app, not a hosted service. Download, paste one soul.md into Manage profiles, connect Telegram (or WhatsApp, Discord, Slack) from the Messaging tab with the same BotFather token and user id, add models under Settings -> Models. No VPS, no gateway token, no terminal. Free — because it runs on your machine and you point it at free open-source models.

Two properties he rates above OpenClaw. First, data location: asked live whether it ships everything to the vendor, the agent answers that conversation history, files, memory and terminal stay local, and only the conversation text and tool results go to the model provider — which is a meaningfully smaller surface than a cloud agent. Second, self-learning skills: under Capabilities you can see skills marked "learned" that the agent wrote itself after doing a task repeatedly. "When you keep doing some task with Hermes it realises this will irritate me, so I'll convert this into a skill." OpenClaw does not do this.

The cost is a hard constraint: the computer must be on. He keeps his PC running; when nobody is home he switches it off and the agent is gone. Hosting Hermes on a VPS is possible but defeats the point, because then it cannot reach your machine.

His argument for a chat app as the control surface is availability. You need nothing installed, nothing configured, and it works from anywhere in the world on a phone you already carry. "You don't need an app, you don't need any setup — you just talk to Telegram and the task is done." That is also why he routes the agent through Telegram rather than the dashboard: the dashboard is for configuration, the chat is for work.

The demos are the evidence, and each one shows a different capability. RESEARCH AND WRITE: an eight-part comparison of AI CRMs — target users, main features, AI features, pricing, best use cases, pros and cons — researched and written into a named Google Sheets tab, in about five minutes against the days or weeks the cohort estimated by hand. BUILD: a weather dashboard web app, built by Hermes driving Claude Code in headless mode, which discovered Claude Code was not installed on his new laptop, installed it via a skill, hit an authentication wall, and kept building anyway — "did it come back saying it is not working, or did it find a way?" CAMPAIGN: one paragraph about a workshop produced a poster with his photo, a QR code, a LinkedIn post and a website. BROWSE AND BUY: "pasta with sauce for four people tonight, check Blinkit, add everything I need to my cart, narrate each step, STOP AT THE CART, do not place the order" — it opened a browser, logged in, worked out the ingredients it was never given, and added six items. COMPUTER USE: Claude took over his screen, opened five PDF invoices in Acrobat and typed the data into an Excel invoice register, then a second agent was asked to verify the register against the invoices and came back asking for approval before continuing.