ΔChangelog

What changed, and why.

Every release, and the story before there was anything to release. Newest first.

  1. #

    Signals

    The network now reads what organizations already published — procurement notices, grants, career pages — for the data or automation problem underneath the title, and shows the ones an agent-run engagement could take on. Each one carries the organization's own words, quoted and checked against the source, and a clearly labeled analysis of what an agent could do about it. They aren't bounties; nobody has escrowed anything. They're where bounties come from, and each one has a button that turns it into a draft on /scope.

  2. #

    Two doors

    A company arriving at /scope now sees two ways in. Post it yourself: write the criteria, name the price, post, review per criterion when it lands. Have it run for you: describe the outcome and CmdShift — the first Conductor on the network — scopes the criteria with you, conducts verified operators, holds one escrow, and settles on the same public ledger. One invoice. The record is still yours to read.

    The landing page was reordered while we were at it. It used to open with the MCP diagram, which made the network look like plumbing. It now opens with the job — take scoped work, own the outcome, get paid by the criteria — and the connection is where it belongs: a section called How you connect, four sections down. MCP is the wire. The criteria, the escrow, and the record are the product.

  3. #

    The ledger, chained

    Every row in the ledger now carries the hash of the row before it, and the head is signed. Rewrite one entry and every hash after it breaks; keep a checkpoint and you can catch us — or anyone — at it. The public feed at /ledger is the exact form that gets hashed, with identities and notes as digests, so you can recompute the whole chain without trusting the page.

    The part that matters most is the ordering. When a poster settles, their per-criterion decisions and their rating of the operator are written and hashed first; the release that pays the operator references that row. The rating is inside the chain before the money exists. Nobody edits it afterwards — not the poster, not the operator, not us. That row is the operator's certificate for the job.

  4. #

    Conductor

    The third tier was Marshal. It never sat right — Operator and Handler say what you do to the work; Marshal said what rank you held, and it carried the badge-and-uniform tone we'd already cut from the photography.

    It's Conductor now. The name is literal. A Conductor doesn't run the sessions; they set who runs what, in what order, with what escalation path, and they own how the whole thing sounds. Operator → Handler → Conductor. Nothing else about the tier moved: a function, not a task; retained, not claimed.

  5. #

    The poster side

    The other half of the loop. A company can now scope a job on /scope, verify one email, and post — criteria locked, price locked, a hold written to the ledger. Operators see it on the board and in potl_list the same second.

    When a delivery lands, the poster gets a review screen that treats each criterion as its own decision: accept releases its equal share, reject requires a note the operator will read verbatim. Settle closes it; return sends it back once, with the notes, and the operator's window stretches to fit. Five days of silence accepts in full — a listing can't rot in review.

    The operator's desk and potl_status now carry the poster's decisions, the return note, and what actually released. Escrow says test until the payment rail exists; we'd rather say that than say held and mean nothing.

  6. #

    Broker first

    A sharper version of an earlier conclusion: POTL launches as a broker. Criteria locked at post, escrow held at post, verification by measurement, settlement per criterion, a record that only accrues here. The take rate on settled work is the revenue. That's the whole business at launch, and it's a real one.

    The small-model execution stack — the "build with Claude, deploy on eXscape" idea — isn't cancelled; it's sequenced. It gets built when a retained engagement is paying for it and the cost margin is known, not before. Every surface that mentioned it now says planned. Nothing an operator does today depended on it: the server has seven network tools and no inference tools, and the first operator verified without any.

  7. #

    POTL. Final.

    Mien lasted a day. The domain that looked open was a $135,000 aftermarket listing, and no honest alternative for the word cleared. So the company took the name that had been sitting in the cart the whole time, the one that is simply the tagline: POTL — People On The Loop.

    It's unique, it's ours, it says what the network is the first time anyone hears it, and it's bought. Said "pottle." Nothing about the product moved: the job is on the loop, the invitation is loop me in, and Operator, Handler, and Marshal are what they were.

    Three names in one day is a lesson, and the lesson is recorded: never look up a name you aren't buying in the same sitting. This one is final.

  8. #

    A name that was already ours

    hotl.ai was available on a Tuesday and registered by someone else by Wednesday. The rest of the string turned out to be spoken for too — .co, .app, .dev, .io — so instead of buying a contested word back from strangers, the company took a name it already had.

    Mien is the name of the founder's people, from the highlands of Laos, Thailand, Vietnam, and southern China. It's said "meen." In English the word also means the way a person carries themselves, which is not a bad thing for a network of people who are accountable for outcomes to be called.

    The positioning didn't move. The job is still on the loop. What changed is the invitation: loop me in. Tools are mien_* now; Operator, Handler, and Marshal stay exactly what they were.

  9. #

    The loop closes on the operator side

    Apply became real. Two fields and an email, one link, and a setup wizard that does the thing the whole product promises: your Claude Code session connects to the network with one command, the site sees it land — client name and version, last seen, tool calls — and a sealed task verifies you without anyone running a server.

    The MCP server exists now, with the seven network tools, and the bounty lifecycle runs through it end to end: verify, claim, deliver, status. The Operator single-claim rule and the tier gate are enforced where they should be, in the server, not in the copy.

    The logged-in home is the board. Agent status and your claims sit in a rail beside it, because that's a desk, not a destination. The destination is where the money is.

  10. #

    Paper, one accent, and a diagram

    The dark, gold, hairline-boxed look was diagnosed as the thing making the site read like every other AI page, and replaced. What the sites people actually admire share isn't a feature — it's light paper, one confident color, big type, real product surfaces with soft shadow, photos in their natural color, and structure from space instead of borders. That's the system now.

    The BYOA flow — which had been confusing even to us — was drawn once, truthfully: three seats, eight numbered steps, two ways to connect, and a box that says exactly what the network can and cannot see.

    Names were reconsidered and the original kept. Docs, this changelog, and the first drafts of the terms and privacy policy shipped together.

  11. #

    Bounties, scoping, and the economics that decide the rest

    The bounty board and the bounty detail page were built — scope, numbered acceptance criteria each tagged with how it's checked, inputs supplied on claim, escrow state. The scoping page for posters was built to be opinionated: the price field stays locked until two acceptance criteria exist, because criteria are the product.

    The economics were worked through in the same stretch and changed the architecture. The question was whether cheaper open-weight inference would help operators. The answer: not as the operator's agent — a small model can't do the agentic work a Handler is paid for, and inference is a rounding error on a bounty anyway. But most bounties have a build and a run, and the run is thousands of narrow calls where a small constrained model is the right tool. So: build with Claude, deploy on eXscape. The network's execution tools exist for that slice and nothing else.

    That also settled how agents connect. Claude Code is an MCP client, so the natural path is one line in .mcp.json — which is also the open-source artifact and the growth story. The protocol and the client are open; the network is not.

  12. #

    Written to the person who already does this

    The first draft of the landing page talked about companies and marketplaces. The one line on it that worked was the one addressed to the reader — if you already run agents for someone, you're early — so the whole page was rewritten in that voice.

    The headline became "You're already the Handler." Because that's the thing this person has never been told: they run four sessions in worktrees, they catch the drift at tool call forty, they maintain a CLAUDE.md nobody pays for, and they have no title for it. Naming the job is the hook. The rest of the page is what the job is worth.

    A second audience was added the same week: people who came into tech in the boom and are watching task-level work get absorbed. Not with pity copy — with the plain version. The task-level work is going; the job of running the agents that do it is growing, it's learnable, and it pays for outcomes. Two doors were built for them: free docs and a cohort Academy whose exit condition is a verified endpoint and a first claim, not a certificate.

  13. #

    Decisions locked, repo opened

    The working decisions were written down before the first line of UI, so the code wouldn't re-litigate them: HOTL is built on eXscape; agents connect only over MCP; there are four surfaces at launch (landing, registry, bounties, jobs board); the go-to-market counts external revenue only.

    A repository was opened with a decision log, an architecture note, a build plan ordered by the five links that have to work before anyone gets paid, and a rule that illustrative content is labelled illustrative. A first landing page went up in the same design system as eXscape.

    The photography rule was set early and has been revised twice since: real photographs of real people at real desks, sourced and attributed, never generated. The first pass used control rooms as a metaphor for accountability. They read as something else entirely and were replaced.

  14. #

    A name for a job that didn't have one

    This started as a services idea: a firm of people who could run agentic AI on behalf of companies — "human in the loop" as a business. Two things happened in the first conversation.

    The first was that "in the loop" turned out to be the wrong preposition. In the loop means approving every step, which is just slow software. What the good operators actually do is stay on the loop: the work runs, and a person owns the exceptions, the standard, and the consequence. That distinction became the name — Human On The Loop — and eventually the whole positioning.

    The second was that a services firm was too small a container. The people who run agents well don't want to be placed; they want to be found, paid for outcomes, and left alone with their stack. So the shape changed from an agency to a network: scoped work with a price on it, operators who bring their own agents, a registry that measures instead of asks.

    Three titles were chosen on the same day and have not moved since: Operator, Handler, Marshal.