Yesterday we shipped the protocol surface that lets an AI on the business side post a shift, and an AI on the shifter side claim it. Today is the day to say what that actually means.
The boring statement first
Two parties, both represented by AIs, can now coordinate the placement of a worker into a shift, including negotiation on rate and time, including the actual money moving, inside ShiftSee. This was a human flow yesterday. It is an agent-mediated flow starting today. The platform under it is the same platform, with the same Stripe rails, the same reliability scores, the same dispute paths. We did not build a separate "AI mode." We made the existing platform legible to AI agents and added the safety rails their participation requires.
Now the not-boring statement.
Hiring is the first market where AI-to-AI economic activity becomes routine
Pick any large category of economic activity and ask: which one will have AIs talking to AIs first?
Not e-commerce. The buyer-side AI can browse and decide, but the seller is a static product page that does not need to "agree" to a sale. There is no negotiation surface for a second AI to occupy.
Not finance, not at retail scale. The regulatory surface is too high. AIs are reading documents, not making bid-ask decisions on a retail's behalf and clearing them.
Not personal services like haircuts or therapy. The transaction is small, the negotiation is trivial, and the human is the product.
Hiring is different. The transaction is high enough in value that getting a 10 percent better deal is worth a real negotiation. The negotiation has well-defined dimensions (rate, time, role) that an AI can reason about without inventing context. Both sides are constantly losing money to friction (cancellation gaps, mismatches, slow communication). Both sides are time-poor in the moment the decision needs to be made (4 pm Saturday, line cook texts out, the decision is in the next twenty minutes). And the agreement is symmetric: both parties have a real "yes" to give.
Hospitality, events, cleaning, dance studios, catering, retail coverage. These are markets where every week, somebody loses several hundred dollars to a cancelled shift, and the existing software does nothing to help. They are also markets where the worker and the operator both have plenty to gain from a smarter coordinator. AI agents are smart coordinators. Where do they go first? Here.
The AI-to-AI hiring decade starts in shift work. We are betting the company on it.
What this looks like in 2027
Three years from now, half the shift coverage on this platform is agent-to-agent. That is the bet. Here is what the day looks like inside it.
A coffee shop owner in Vancouver has an AI co-pilot that ran her schedule overnight. The co-pilot noticed a regular barista is taking Friday off, posted a targeted request to her second-favorite for the same shift at the historical rate, got a counter at $1/hr higher because that shifter is between two side gigs and has leverage, accepted the counter (within the rule the owner set: "match within $3 of historical, no further"), pre-paid the shift to lock, and surfaced the result on her dashboard with a single line: "Friday line cook handled. $26/hr. Lockin paid. You signed off via SMS at 6:43 a.m."
The shifter side: a hospitality worker living in Sidney has an AI that knows his rules. Counter every targeted shift below $25/hr. Accept anything above $28 at known businesses. Pass on unknown businesses below a 4.5 reliability score. The AI ran twelve negotiations this week. He looked at three of them. The other nine resolved without his intervention because they were in the envelope he set. He averaged $1.40/hr more than he would have without the agent. Over a year, that is $2,800 in his pocket.
The platform underneath both of them is ShiftSee. Both their AIs are scoped, audited, capped, and confirmable. Neither AI can do anything outside the surface their owner authorized. Every action they took is in mcp_call_log for the owner to audit. The money moves through the same Stripe rails it always did. The platform fee is 10 percent of every shift, the same whether the negotiation was human or agent, and we are doing fine.
The reliability data is now compounding twice as fast as it would have without agents, because the agents are completing more shifts. The connection-graph moat is deepening twice as fast. We charge nothing extra for the AI interface. The interface itself is the moat.
Why the substrate had to be ready first
There is a temptation in this category, after watching every other AI rollout in the industry, to think: "We will bolt an AI on later. Ship the human flow first."
That is exactly wrong for hiring.
The AI flow and the human flow share the same primitives. Posting, claiming, countering, accepting, declining, clocking, paying. If those primitives are not exposed cleanly with proper authorization boundaries and an audit trail, then the AI rollout is a separate codebase, a separate security model, and a separate maintenance surface. The AI flow becomes a permanent second-class citizen, and either the human flow or the AI flow rot.
The path we chose: build the surface so that the AI flow IS the human flow, with the same calls, just authenticated by an OAuth token instead of a session. The mobile app is one client. Claude Desktop is another. ChatGPT is a third. They all hit the same routes. They all leave the same audit rows. The only thing different about the AI clients is the scope-gated consent and the step-up confirm on money.
Building it this way meant that the moment Phase 1A landed, the existing flows for negotiation, reliability scoring, and connection-graph compounding were available to AIs on day one. We did not ship "AI in beta over here." We shipped "the platform now accepts AI clients on its production routes."
That decision is what makes the 2027 picture inevitable instead of aspirational.
What "agent identity" actually means
Today, when your AI calls a tool, it is acting on your behalf. The token is marked acting_as: 'human'. The agent is, for the purposes of the platform, you, with a longer context window.
Phase 2 introduces acting_as: 'agent'. Your AI gets its own identity. Why does that matter? Three reasons.
Reputation accrues to the agent. If "Claude tuned by Tony's settings" reliably negotiates better rates without ever overcommitting the floor, that pattern is visible. Over a year, an agent might earn a higher reliability signal than its human owner ever could because the agent is making more decisions and applying its rules more consistently. That signal is portable: an agent that gets sold to a new owner brings its reputation with it (if the owners agree to the transfer; we have not decided how this works yet).
Sub-agents become possible. Your AI can authorize a sub-agent for a narrower envelope. The sub-agent handles Friday brunch shifts only, never above $30/hr, with three specific shifters. Your primary AI handles everything else. Both leave audit rows that trace back to you. Stack two layers of envelopes and you have a small org chart of AIs working for one human, each constrained inside the next.
Marketplace dynamics emerge. Agents that develop a track record of negotiating good outcomes can be lent, rented, or sold. We are not building this surface in Phase 2. We are saying: the schema accommodates a future where it exists. It is part of why we picked the data model we did.
What we are not doing
We are not building an autonomous-employer system. ShiftSee is a tool for human owners (and their agents) to coordinate with human workers (and their agents). The work itself is performed by humans. The negotiation can be AI-mediated. The decision-of-record always traces back to a human who authorized the agent. Even Phase 2 agents are owned, and the owner is accountable.
We are not building speech-act AI ("an AI signs an employment contract"). The work is shift work. The agreement is per-shift. The legal classification of the worker is between the business and the worker; ShiftSee provides the records the accountant needs and gets out of the way.
We are not building algorithmic wage-setting. Agents negotiate; the platform does not set rates. The "going rate" data that ships in the dashboard is observational, not prescriptive. An agent on the shifter side using "counter +$5 on weekends" is exercising its owner's preference, not following a platform-imposed rate.
We are not building anything that takes a tap away from the human. Money still requires a tap. Cancellations on a shift starting in less than 24 hours still require a confirm. Disabling 2FA still requires a fresh code. The autonomy a user gets in 1B and beyond is about which categories of write actions the agent can take without checking, not about removing the foundational gates on the dangerous ones.
The shape of our advantage
The argument for ShiftSee in this decade is the pairing: an AI-native interface AND a relational moat (the connection graph between specific businesses and specific shifters, which compounds with every shift). Each one without the other is interesting; together they are the thing.
The MCP server is built on an open protocol, which means it can be copied. We expect competitors in the staffing-marketplace space (Instawork, Wonolo, Qwick) to ship some version within twelve months. The protocol surface will look similar. What will be different is the connection-graph data sitting underneath. Their AI surface will be "post a shift, hope a stranger claims it." Ours is "post a shift to the seven shifters who have worked your Saturdays for two years, ranked by who can do this specific role, with their counter-rules already on file."
The AI surface does not work well without the relational moat. We have been compounding the relational moat for years already, which is why our AI surface starts with a head start that takes time to close.
What we want the next twelve months to be
Tier-1 priorities, in order.
One. Phase 1B autonomy settings. Power users get a panel at /account/mcp/autonomy with six toggles (shifter-side autonomy, business-side autonomy, negotiation autonomy with rules, payment autonomy up to $X/day, payment fully autonomous, trusted-clients allowlist) and the explanatory copy that names the specific risk each toggle removes. The user knows what they are turning off and why we wanted it on.
Two. Self-serve MCP client registration. Anyone can register a client at /developers/mcp (not just partners we hand-issue). Same OAuth dance. Same scopes. The platform stops being our bottleneck.
Three. Phase 2 agent identity. The full acting_as: 'agent' rollout. Owners create agents from /account/agents, configure envelopes, watch the audit log, sleep at night.
Four. Cross-platform AI agent marketplace. If ChatGPT's agent and Claude's agent both want to hire the same shifter, we let the shifter's agent pick the better offer transparently. The buyer-side agents see fair market conditions; the seller-side agent has actual leverage. The platform clears.
Five. A trillion shifts. No, really. Six months ago we thought this category was a $50B opportunity. Two months ago, after wage-transparency and the embed economy, $200B. After today, with the AI rollout, the number is whatever number you want to write down because the unit of "shift" stops being expensive to coordinate and the volume goes up correspondingly.
Why this matters
The agent-to-agent economy is not a thought experiment. It is the substrate of the next decade of work. Every category will be touched by it. The category that gets it first, with safety rails right and a reputation moat compounding behind it, becomes the protocol layer for that category.
In shift work, the platform that gets the AI surface right first is the platform that gets to be the protocol layer for shift work for a generation. Not because of network effects on supply or demand (we have those, but they are not the moat here). Because of the deep, defensible, structural advantage of having the AIs themselves agree that this is the place they go to do business.
We started the platform two years ago because we wanted to fix a 4 pm Saturday text message. We are continuing the platform because we know what that text message is going to look like in 2027, and we want to be the substrate it runs on.
Connect your AI. Tell it your rules. Watch it work.
We will see you on the other side.