The Multiplayer Surface Is Free. Coordination Capital Is Not

In June a frontier lab shipped the shared agent in beta. By 31 July it was free under an MIT licence, and none of the five vendors tracked here had published a price for it.

On 31 July 2026 Y Combinator open-sourced the agent harness behind its own accounting, legal, events and engineering. Its repository opens: "A multiplayer agent harness for work. In Slack and on the web." The licence is MIT and the price is zero ^[1]. Thirty-eight days earlier, on 23 June 2026, Anthropic had put a single Claude inside selected Slack channels and named the arrangement in its own words: "@Claude is multiplayer. Within a given Slack channel, there's one Claude that interacts with everyone" ^[2]. That shipped on two paid plans, on one surface, and was still in public beta on 18 August 2026 ^[3]. The caveat belongs in the same breath: the open release is version 0.1.x, its authors call it early and buggy, and stars are not deployments ^[1]. It is still the hardest fact here, because it prices the capability instead of rating it.

So this thesis is not about either product but about a price that does not exist. On 18 August 2026 I read the published pricing pages of the five vendors that market a shared agent living in a team surface — Anthropic, OpenAI, Microsoft, Google and Asana. Every one of them publishes a seat price. None of them publishes a price for the shared agent itself ^[3][4][5][6]. That is a census of five pages on one date, not a record of what buyers sign. The capability every vendor now markets is the one thing none of them sells, and a party outside the vendor set has already established what it fetches.

Executive synthesis

The channel is where coordination happens. It is not where coordination accumulates.

Everyone building on frontier models lives inside one tradeoff. Option A is to stay in the subscription harness — the seat-priced products the labs sell directly — and build workflows, bundles and whole operating systems inside someone else's product. Option B is to build an application or an agent against the API. The observation that triggered this thesis is that Option A had no multiplayer mode and that Claude Tag lands exactly in that gap. Section 03 shows it true of Anthropic and false of the industry; the question underneath survives the correction. Where does coordination capital — canonical objects, the roles and permissions attached to them, approval chains, decision history, and the accepted definition of when work is done — accumulate once the harness goes multiplayer, and to whom does the margin flow? My answer: the surface commoditizes on a schedule you can already read off the record, the state underneath it does not, and today almost none of that state belongs to the lab.

• 0 of 5 Tracked vendors publishing a price for the shared-agent surface as its own product (Anthropic, OpenAI, Microsoft, Google, Asana) Vendor pricing pages and documentation read 18 Aug 2026 MEASURED absence across five pages
• $0 per seat Claude Tag's added seat charge. Channel work bills to an organization usage balance under a spend cap; no per-unit rate is published Anthropic documentation, read 18 Aug 2026 · product announced 23 Jun 2026 COMPANY CLAIM
• 30M+ Paid Microsoft 365 Copilot seats, with net seat adds more than doubling quarter over quarter, and a model described as moving "beyond per-seat to per-seat-plus-consumption" Microsoft FY2026 Q4 earnings call, 29 Jul 2026 COMPANY CLAIM

Reading note, including the part that cuts against me. The third card is the strongest counter-evidence in this piece and it sits in the synthesis on purpose ^[7]. The seat is not dying; it is being re-based — priced for access and governance, with meters accumulating beside it rather than replacing it. Everything below has to survive that sentence. And no agent-behaviour claim here has been verified outside the vendor selling it.

What changed in June 2026 was not sharing. It was an agent with its own identity and its own budget.

Anthropic's head of product for Claude Code and Cowork, Cat Wu, told Fortune on launch day: "Claude Code, Cowork, and chat are very single-player, whereas Claude Tag is built to be interactive and multiplayer" ^[10]. Anthropic's own documentation says the same of the product Claude Tag replaced ^[11]. For Anthropic's harness the observation is not merely defensible — it is the vendor's own position, stated against interest.

For the industry it is wrong. OpenAI shipped shared agents on 22 April 2026, built once and used together "in ChatGPT or Slack" ^[6]; Microsoft had Copilot in Teams group chats generally available in February 2025, sixteen months earlier and outside this window ^[7]. The narrow claim that survives: until 23 June 2026 no harness had shipped an agent acting under its own admin-provisioned service account rather than as the person who asked, with tool scope and a spend cap attached to the channel ^[3]. That is authority attached to a scope — smaller and sharper than "multiplayer".

And one claim of my own, withdrawn in public. My first pass through this record called Claude Tag the first ambient example. It is not. The term belongs to Harrison Chase of LangChain, published 14 January 2025, before this window: an ambient agent listens to "an event stream and act[s] on it accordingly" ^[12]. Against that test Microsoft's Facilitator agent was generally available in Teams chats in April 2025, following the conversation without per-turn invocation — fourteen months earlier, also outside the window ^[7]. Inside the window, OpenAI's launch text says its product team built an agent that "proactively answers employee questions in Slack channels", two months before Claude Tag ^[6]. Anthropic does not claim the category either: it uses "ambient" once, in quotation marks, for a behaviour an administrator can switch off ^[2]. And the narrower claim that replaces mine is itself twenty-eight days old: Block's Buzz, released 21 July 2026 under an Apache licence, gives every agent its own keypair and audit trail ^[1].

The platform owner said the quiet part on the record. Rob Seaman, general manager of Slack, told Reuters at launch: "A lot of the capabilities did exist, but actually the form factor of being able to tag it the same way that you would a coworker is really powerful" ^[13]. The company with most to gain from calling the surface an asset called it a form factor.

The chronology behind the false-scarcity call runs one way, with a single exception. Asana used "multiplayer mode" in its own release of 4 June 2026, nineteen days before Anthropic used the word ^[9]; Notion, Salesforce and Google each put an agent into a shared surface of their own, the last two deliberately personal ^[14]. The exception matters more than the rows that agree: OpenAI began winding down ChatGPT group chats on 9 July 2026, sixteen days after Claude Tag ^[6]. Nobody defends a moat by retiring it weeks after a competitor walks in.

The meter the industry actually runs is not the one the sales narrative describes.

The mechanism: scarcity moved to authority, and authority is not in the channel

The enabling change is not that models got good enough to sit in a channel; it is that the price of a good answer keeps falling. Anthropic publishes Claude Opus 5 at $5 per million input tokens and $25 per million output, against $15 and $75 for the model it replaced, though a roughly 30% higher tokenization from Claude 4.7 onward makes the real cut nearer 2.3-fold. It also cancelled a pre-announced 50% increase on its workhorse model thirteen days before it was due ^[4]. Whatever is scarce, it is not the answer.

The amplification is the shape of the shared surface. One agent's output is consumed by twenty humans; one human's instruction produces thousands of machine actions. Both ends break the seat: it over-counts the consumers of the work and under-counts the work. The inverse shape shipped a week before this cutoff — xAI's Grok Bot, whose group chats coordinate one account holder's own agents rather than a team's people ^[17]. The constraint that then bites is not capability but authority. A message in a channel makes intent visible; it does not establish that the actor was permitted to act, that the object it touched was canonical, that a budget covered it, or that anyone can prove afterwards what happened — each of them a property of the organization rather than of the model.

The scarcities, ranked, because they are not equals. First, authority that travels — an identity, a represented principal, a mandate, a budget and a policy that survive a system boundary. Nothing else can be automated until this is solved, and no normative profile has shipped ^[7][3]. Second, canonical state with write rights — an agent can read almost anything; it can only finish work where it may write. Third, an accepted definition of done, without which there is neither outcome pricing nor exception accounting. Fourth, accumulated conventions and decision history, the stickiest form of coordination capital and the hardest to defend because it is invisible in financial statements. Fifth, presence in the shared surface, which is real and is rented.

And one false scarcity: the shared surface itself. Six vendors reached the same product shape in nine months, and nobody charges for it. Then the argument stopped needing a forecast: within ten days that July, a payments company and an accelerator released the human-shared version at zero ^[1]. My reading is that "an agent everyone in the room can see" becomes table stakes across the harness layer within roughly eighteen months, and that treating it as an asset is the most expensive mistake available to a founder in 2026.

The new unit of value is the workflow instance — an end-to-end run, auditable, with a cost and an outcome attached. The cleanest published artefact of that is Anthropic's billing schema, which charges shared work to the organization and private work to the individual's seat ^[3]. The cost boundary was drawn exactly along the coordination boundary.

Counterforces, at full weight. The organizations supposed to own well-formed coordination capital largely cannot describe it: in three surveys published on 29 July 2026, 71% of developers said their agents can reach sensitive information and an aggregated governance maturity score came in at 35 out of 100 ^[18]. Every publisher in that stack sells into the gap it measures, so treat the levels as directional. The edge that cuts: an asset nobody can enumerate is a poor moat, and it hands the advantage to whichever vendor arrives with ready-made governance. Add instruction conflicts and credit meters that destroy budget predictability.

The data side of coordination capital now has a public price tag: Google won a bankruptcy auction at a press-reported roughly $10 million for the corporate corpus of the liquidated Spirit Airlines, including around 500 million Microsoft Teams messages — announced, and pending court approval as of 18 August 2026 ^[19].

Value map: the surface is claimed, the authority is contested, the evidence layer is empty

Five layers sit between an intention typed into a channel and a change made to a record. By 18 August 2026 they are not equally available.

• Shared surfaceCLAIMED

  The room the agent lives in: channel, board, group chat, meeting

  owned by Slack, Microsoft, Google and Notion; rented to everyone else, including the labs ^[14]

• Agent identity and budget scopeCONTESTED

  Which principal the agent represents, what it may reach, what it may spend

  shipping proprietarily on both sides — service accounts per channel at Anthropic, directory identity and entitlement governance at Microsoft — with no normative profile in sight ^[3][7]

• Semantic context over workCONTESTED

  What this document, message, project and person mean to each other

  the incumbent suite is building it as first-party product and exposing it through open protocols ^[7]

• Canonical objects with write authorityOPEN

  The customer, the deal, the case, the invoice — and the right to change them

  still the customer's, reached through connectors; no frontier lab has shipped a record of its own ^[11]

• Evidence of what the agent didOPEN

  Which writes an agent made, under whose authority, at what cost, reversible how

  the least-covered layer: no public series measures agent-initiated writes to canonical systems

The bottom row is the one I would spend a career on. This thesis's strongest falsifier — whether shared agents ever move beyond summarizing and suggesting into authorized writes — cannot be measured from any public source, and saying so beats implying otherwise.

The room, its memory and the assistant in it are a feature waiting for a product manager at Slack, Microsoft or Google to notice. Salesforce made the point on 13 January 2026, shipping its own agent into its own surface five months before its most prominent tenant did ^[14].

Where the economics work: four categories with real meters, and the risk inside each

These are not "AI sectors" but the places where coordination capital is already being paid for. Companies named are publicly documented examples, not endorsements; I hold no position in, and no client relationship with, any.

And one category with no meter at all. Around the subscription harness a market has grown of packaged systems that turn a flat plan into an operating system for a function or a whole company, sold through creator-commerce platforms and directories. This is Option A's native economy: nobody in it calls an API or operates a system of record. I name no seller — the finding is the category's price structure, not anyone's package. The largest third-party directory I could read advertises more than 800 packages and twenty curated bundles, free; paid listings cluster between roughly $19 and $167 ^[21]; and the lab that owns the harness shipped a fifteen-workflow small-business bundle of its own on 13 May 2026, with no price stated anywhere in the announcement ^[11]. What is sold is procedural knowledge, priced like content rather than software, with nothing accumulating on the seller's side: conventions without canonical state, the fourth scarcity stripped of the second. No platform publishes churn or retention data for the category, so nobody can honestly say more.

Which points at where multiplayer actually lives. The natural habitat of a shared agent is the software that already holds the multi-party object — the deal, the case, the invoice, the project. Signal R02.5 records the baseline: no lab has shipped a first-party store of canonical objects, and when one reached for small-business coordination in May 2026 it shipped connectors into other companies' records instead ^[11], routed through Cowork — the surface Anthropic's own head of product would call "very single-player" forty-one days later ^[10]. A company operating system with no mode for the company to be in it together. The pattern is not the lab's alone. On 5 August 2026 Cloudflare open-sourced the company workspace it had run internally since May: every shared app is copied from a blueprint into a separate instance with its own database, a user's agent is scoped to that user's existing permissions, and staff had built more than 4,000 apps in thirty days COMPANY CLAIM ^[22]. Everyone ships around the canonical store; nobody ships it. Cloudflare is not a frontier lab, so R02.5 does not move — and the vendor giving the harness away sells the substrate it runs on. Those connectors reach collaboration only by riding systems already built for it, and what they cannot borrow is a shared agent: each session runs under one person's account ^[11], so the team sees the output in the record and never the reasoning in the room. Rented state, rented collaboration, and the coordination capital accrues to the landlord. The open-source entrants do not change that: one owns the relay, one the scoped memory, one the loop; none owns the object ^[1]. What makes that object durable is a different argument, reserved for a later thesis.

Market structure: eleven months, six vendors, one retreat, two giveaways

Product status matters more than announcement volume here, because a beta on two plans on one surface is a narrower event than a launch post implies, and every row states what had actually shipped. The last row is here because it does not sit on the same axis: counting a personal agent as a multiplayer event is the commonest error in this category.

1. 25 Sep 2025

   Asana announced AI Teammates — agents that live in the project rather than in a private chat, with one agent able to support several teams at once ^[9]

   Announced — betaFirst shared agent in the windowCOMPANY CLAIM

2. 13 Nov 2025

   OpenAI piloted group chats in ChatGPT for 1–20 people on consumer plans, with personal memory deliberately excluded from the group; expanded globally on 20 November ^[6]

   PilotShared conversation, not a shared agentCOMPANY CLAIM

3. 18 Nov 2025

   Microsoft announced Teams Mode for Microsoft 365 Copilot, agents in Teams channels working with third-party agents over the Model Context Protocol, and Agent 365, described as "the control plane for agents" ^[7]

   Public previewAuthority layer, first-partyCOMPANY CLAIM

4. 8 Dec 2025

   Anthropic shipped Claude Code in Slack: mentioning @Claude in a channel started a coding session that ran "under an individual user's account", on that user's repositories and plan limits ^[11]

   BetaSingle-player in a shared roomCOMPANY CLAIM

5. 13 Jan 2026

   Salesforce brought a rebuilt Slackbot to general availability, positioned explicitly as "your personal agent for work" — the surface owner occupying its own surface, with a single-player product ^[14]

   Generally availableLandlord moves inCOMPANY CLAIM

6. 22 Apr 2026

   OpenAI announced workspace agents: shared agents built once and used together "in ChatGPT or Slack", with memory, schedules, approvals and admin controls; free until 6 May 2026, then credit-priced with no published per-unit rate. Google announced a Gemini agent in Google Chat the same day, framed as a personal assistant ^[6][14]

   Research preview at launch; generally available by 18 Aug 2026, no dated announcement foundShared agent, two months earlyCOMPANY CLAIM

7. 13 May 2026

   Anthropic announced Claude for Small Business: 15 agentic workflows and 15 skills running through Claude Cowork, connecting to QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace and Microsoft 365. No price stated in the announcement ^[11]

   AnnouncedThe lab rents canonical state rather than building itCOMPANY CLAIM

8. 16 Jun 2026

   Microsoft's Work IQ interfaces reached general availability — a semantic layer over organizational work, consumption-priced in credits, reachable over Agent2Agent and a remote Model Context Protocol endpoint, with data and actions staying inside the tenant boundary ^[7]

   Generally availableThe incumbent builds the underneathCOMPANY CLAIM

9. 23 Jun 2026

   Anthropic launched Claude Tag: one @Claude per Slack channel, admin-provisioned service-account identity, per-channel tool scope, channel and workspace memory, organization and per-channel spend caps, scheduled work over "hours or days" ^[2][3]

   Public beta — Team and Enterprise plans, Slack only, still beta on 18 Aug 2026Authority attached to a scopeCOMPANY CLAIM

10. 1 Jul 2026

    Notion released external agents in general availability: agents from other vendors assignable and mentionable from a team-shared board, following an alpha on 13 May 2026 ^[14]

    Generally availableCross-vendor invocationCOMPANY CLAIM

11. 9 Jul 2026

    OpenAI began winding down ChatGPT group chats — no new group chats, existing ones read-only — sixteen days after Claude Tag shipped, with the stated reason that it was simplifying the experience ^[6]

    RetirementThe one row that runs backwardsCOMPANY CLAIM

12. 21 Jul 2026

    Block released Buzz under an Apache licence: humans and agents in the same channels, with each agent holding its own keypair and audit trail — "Agents are members, not bots" ^[1]

    Released — open sourceThe scoped identity, given awayCOMPANY CLAIM

13. 29 Jul 2026

    Microsoft reported more than 30 million paid Microsoft 365 Copilot seats with net seat adds more than doubling quarter over quarter, and described the business model as moving "beyond per-seat to per-seat-plus-consumption" ^[7]

    Reported — regulated disclosure venueAdditive, not substitutiveCOMPANY CLAIM

14. 31 Jul 2026

    Y Combinator open-sourced QM under an MIT licence — "A multiplayer agent harness for work. In Slack and on the web" — with per-person and per-room scoped memory, permissions, sandboxes and schedules, run internally across accounting, legal, events and engineering ^[1]

    Released — open source, version 0.1.x, described by its authors as early and buggyThe surface at licence price zeroCOMPANY CLAIM

15. 11 Aug 2026

    xAI opened an early beta of Grok Bot: a standalone personal application, one cloud computer per account, whose group chats coordinate the account holder's own agents rather than a team's people ^[17]. Nous Research bundled multi-agent rooms into its own free harness on 18 August 2026 ^[1]

    Early beta — vendor primary not opened; reported through trade pressThe other axis: one human, many agentsCOMPANY CLAIM

Four paths, and the one I hold

Probabilities are my judgement, stated so they can be scored later; each path names one leading indicator, mapped to a signal in section 10.

Base — parity arrives and is priced beside the seat (45%). The shared agent reaches general availability across most harnesses, stays bundled or metered from a pooled balance, and the seat persists as the access-and-governance charge. I hold this path: it is the only one consistent with all four things I can observe at once — fast convergence on the shape, no price on it, seats growing at the largest vendor, and no lab building a record of its own. Indicator: R02.3 stays at its confirming condition while R02.1 climbs toward four of five.

The suite annexes the underneath (25%). Microsoft, Google and Salesforce ship semantics, agent identity, entitlement and audit as first-party product inside their own tenants; the labs remain tenants with excellent models and thin governance. This is the best-supported alternative — control plane, semantic layer and directory identity all shipped from one vendor inside this window. Indicator: R02.4 weakens as surfaces close to first-party agents, while R01.4 stays neutral or falls.

The lab annexes canonical state (15%). A frontier lab ships a first-party store of business objects with write authority, roles and an audit trail, and accumulates coordination capital rather than borrowing it. The one dated piece of evidence points the other way ^[11]. Indicator: R02.5 flips to its confirming condition.

Null — shared agents stay assistants (15%). Channel agents remain confined to summarizing and suggesting; authorized writes to canonical systems do not rise; the meters stay small and the seat absorbs everything. Here the migration was cosmetic and this thesis is wrong at the root rather than at the edges. It is also the path with the weakest instrumentation. Indicator: R02.2 stays at −1 with no cohort member disaggregating a non-seat line, while R02.1 stalls at two or fewer.

What others say, and where I differ

SemiAnalysis, "AI Value Capture" (1 May 2026), is the strongest thing written against this thesis: value capture across the stack moved to the model labs during 2025 and 2026 ^[23]. It is a modelled attribution behind a paywall, used at the level of its claim rather than its figures. My departure is about duration, not direction: lab margin is capability rent collected during a race in which answers still differentiate, and falling list prices erode it. If lab margin is still climbing in 2028 while no cohort vendor discloses a non-seat line, I was wrong and the rent was the business.

Bessemer's "systems of action" (28 May 2025, prior art from before this window) and Foundation Capital's services-as-software work (18 August 2026) both move the prize above the record: agents absorb the labour surrounding a system of record, and, in Foundation Capital's sharpest line, integration "is the product surface" ^[24]. I converge on the mechanism and depart on the owner. Their version accumulates encoded domain knowledge inside the AI-native vendor; my reading is that the customer already owns most of it, and the vendor's problem is borrowing write authority it will never own. Sarah Wang, in the same a16z collection as the claim adjudicated above, takes the opposite side of my second-ranked scarcity: the system of record becomes a commodity persistence tier while advantage moves to whoever controls the execution environment ^[8]. Her reading holds if agents can freely write to the record; mine holds if write authority stays scarce and slow.

The vocabulary is older than the category. Malone, Yates and Benjamin argued in 1987 that cheaper coordination would shift activity toward markets rather than hierarchies; Malone and Crowston later defined coordination as the management of dependencies among activities ^[25]. Their prediction runs opposite to mine, which is why naming it beats borrowing it quietly: my claim is that the residue — authority, canonical state, a definition of done — does not get cheap, and therefore concentrates. The organization-capital literature behind the term has spent forty years unable to agree on a measurement, which is why the fourth-ranked scarcity is the hardest to sell to an investment committee.

Signal scorecard

Five new signals join the public ledger with this thesis, and three from Thesis 01 take new readings — one of them a re-score. Scores are mine, against each signal's confirming condition. Read the sign carefully on R02.5: a negative score there means annexation is not happening, which supports the argument. The full ledger is at radar.ersiner.ai/signals.

The re-score is the one to argue with. Identity incumbents have reached general availability, which is half of R01.4's confirming test; no normative profile stands behind them, which is its weakening test. Both are true at once, so neutral is the honest reading rather than a drift.

CALL A — by 31 December 2027, at least four of the five named harnesses (Anthropic, OpenAI, Google, Microsoft, Meta) will have a generally available shared agent living in a team surface with persistent group state and its own scoped identity. Confidence 75%. Ledger signal R02.1. The baseline makes it demanding: on the strict reading the count today is zero, because Anthropic has not cleared its own bar. Two things cut against me — parity in collaboration features has historically taken years rather than quarters, and the convergence was not independent, since a widely read investor essay told the category to build exactly this in December 2025 ^[8]. July's open licences do not resolve it; none of those parties is in the cohort.

CALL B — through 31 December 2027, no member of the fixed cohort (Salesforce, HubSpot, Atlassian, monday.com, Asana, Smartsheet) will report a separately disclosed non-seat agent revenue line exceeding 10% of total revenue. Confidence 65%. Ledger signal R02.2. The rationale is not that vendors are silent, which would be false ^[20]. It is that disclosure exists and scale does not: $1.2 billion against roughly $44 billion annualized is about 2.7%. This is a bet on a reporting convention as much as on economics.

CALL C — by 30 June 2028, no frontier lab will have shipped a first-party, general-purpose store of canonical business objects of customer-record or resource-planning class as a product. Confidence 70%. Ledger signal R02.5. The test of whether the harness annexes coordination capital or keeps renting it.

Regional translation: the seat is priced globally, the coordination is priced locally

The harness sells one number to the whole world. A standard team seat costs the same in Bucharest, Dubai and San Francisco, and so does a million tokens ^[4]. What the seat displaces does not. Take a coordinator whose fully loaded monthly cost is $6,000 in a high-wage market and $1,500 in a lower-wage one. The $20 seat is 0.3% of the first and 1.3% of the second: trivial in both, and therefore not the decision. The meter is. At the run costs in section 04 one busy channel costs $4,200 a month in both markets — 0.7 of a coordinator in the first and 2.8 in the second INFERENCE arithmetic on published rates and stated assumptions. Automation that is obviously cheap in one market is a real cost decision in the other, and flat global pricing does the deciding. Geography adds a line item on top: a 1.1× multiplier for inference restricted to the United States, a 10% premium on regional endpoints at two cloud marketplaces ^[4].

The regional argument for Central and Eastern Europe and the Gulf is one line: a business spanning several regulatory perimeters holds coordination capital worth more, not less, precisely because it is unpleasant — an approval chain satisfying two regulators at once has an owner, a clock and an audit trail, and a general agent must borrow it rather than infer it. The exportable asset is not cheaper engineering; it is the governed process itself.

Risk: the wage figures are illustrative assumptions, not a measured series, and the ratio moves with the exchange rate first.

Stakeholder lenses

Investors — the buyer is an incumbent defending a control point.

Ask what the company owns that the surface owner cannot ship in a release note. Then ask for the write list: which canonical objects the product may change, under whose authority, and what the audit trail shows when it is wrong. The useful diligence artefact is not the demonstration, it is the permission model. The same test sets the exit: surfaces are cheap to build and expensive to defend, records are the reverse — expect an incumbent protecting a control point, not a platform buying growth. And underwrite the shared surface as rent: nobody in the tracked set sells it as its own product, and two outside it released it at zero ^{[1] [3]}.

Founders — the undifferentiated middle is gone; pick a side on purpose.

Option A keeps strengthening: the harness has multiplayer, workflow templates and connectors into other people's records, priced so the seat is never the objection. Option B is not dead, but a single-model wrapper with no state of its own is squeezed from both sides. What survives on either side is an object model, a permission model, and a definition of done the customer agrees with. I write from inside this tradeoff: the system behind this series runs inside a subscription harness, not against the API.

Enterprises — a shared agent is a spending decision wearing a chat interface.

Fix three numbers before the pilot: the spend cap per channel, the systems the agent may write to, and the definition of a finished task you, not the vendor, will judge by. Date every adoption figure collected before 1 September 2026 against the launch credits funding it: $25,000 per Enterprise organization and $2,500 per qualifying Team organization, expiring that day ^[3]. Your side of the table holds the write path, and it now carries a public number: a press-reported roughly $10 million for a liquidated airline's corporate corpus, 500 million Teams messages, pending court approval ^[19]. The counterforce is internal: 71% of developers say their agents can reach sensitive information, governance maturity 35 out of 100 ^[18] — an asset nobody can enumerate is a poor moat.

Careers — permission work appreciates, coordination work gets metered.

Both sides are here for anyone advancing or changing careers. What appreciates turns organizational practice into something a machine may act on: permission design, agent operations, spend governance, naming which object is the truth, and section 05's empty row — which writes an agent made, under whose authority. The exposure is the same arithmetic. The meter is quoted against a coordinator, not a seat: one busy channel at $4,200 a month is 0.7 of a coordinator in a high-wage market and 2.8 in a lower-wage one INFERENCE — convenience in one market, substitution in the other. Outcome pricing moves the billing unit off the person: $0.50 per resolved conversation, $1.00 per recommended lead ^[20], against a definition of done the vendor writes.

Method, sources and disclosure

Evidence window: 1 September 2025 to 18 August 2026. Every figure was re-checked against the linked primary source on 18 August 2026. Items dated before the window — Microsoft Copilot in Teams group chats (February 2025) and its Facilitator agent (April 2025), the 2025 restriction on third-party retention of Slack conversational data, the ambient-agent definition (January 2025) and Bessemer's roadmap (May 2025) — are carried as historical precedent, never as evidence about the window. Publication date, event date and measurement period are kept separate: the pricing and margin figures come from surveys fielded in April and December 2025 and published in January 2026; the monetization figures from a survey fielded in April and May 2026.

What could not be closed. Several sources block automated fetching and were read at headline level only, including the value-capture analysis in section 09. The vendor page for Grok Bot could not be opened at all, so every fact about it here is trade-press reproduction of launch copy; the accelerator's launch post was read through search indexing and corroborated against its repository. OpenAI's pricing pages and Google's US-dollar list prices could not be read from this workstation, which is why the comparison in section 01 prints no OpenAI number and describes Google's structure rather than its price. In-window filings for two of the six R02.2 cohort companies were not opened, so that baseline reads four of six.

Incentives, named. Every capability statement here comes from a company selling or promoting it — Anthropic, OpenAI, Microsoft, Google, Asana, Notion, Salesforce, Block, Cloudflare, Nous Research, xAI and Y Combinator describing their own products. The margin and pricing surveys come from a growth investor in the companies it surveys, whose deck states the sample includes some of its own portfolio; the monetization survey from a pricing advisory practice, on a self-selected sample. The adjudicated forecast is published by a venture firm invested in the category it advantages, as are both roadmaps in section 09; the governance surveys come from three vendors selling into the gap they measure; the ambient-agent definition from a vendor selling the tooling that category requires; and an accelerator open-sourcing the harness its portfolio would otherwise buy is doing marketing as well as engineering.

Evidence classes, one per stat card, chart and timeline entry: MEASURED an observed count or survey result with its population stated; COMPANY CLAIM a figure a company publishes about itself; FORECAST a named forecaster's projection; INFERENCE my own arithmetic or judgement, inputs shown.

Vocabulary, attributed. "Multiplayer" is not my category: it is quoted from the named people and companies who used it — an investor in December 2025, a chief product officer in April 2026, a vendor in June 2026, an accelerator in July 2026. "Ambient agent" is Harrison Chase's, published by LangChain on 14 January 2025; where this piece tests the word it tests his definition ^[12]. "Coordination costs" and coordination-as-dependency-management belong to Malone, Yates and Benjamin (1987) and Malone and Crowston (1994); "organization capital" belongs to the intangible-capital literature, and coordination capital as used here is that concept narrowed to its machine-readable, permission-bearing subset ^[25]. "Control plane" is a borrowed analyst category. The ranking, the false-scarcity call and the scorecard are my own.

How this was made. This thesis began as my own draft position, and every claim was treated as something to test. Three lost: the harness did have a multiplayer mode before June 2026; seat growth is not visibly decoupling from value; and the trigger product was not the first ambient agent. All three are argued with in the text rather than deleted.

Disclosure. Written in a personal capacity, from public sources only. The author works within the Türkiye venture ecosystem; to avoid conflicts of interest, no Türkiye-based fund or startup is named or evaluated here. The author holds no positions in, and has no client relationship with, any company named.

This is Thesis 02

Second thesis in a recurring series. Theses are numbered across topics and published when ready, never on a calendar; each carries its own signal scorecard, and a bi-monthly Scorecard re-scores every live signal no publication has touched and grades the dated calls in public — including the misses, which are the only ones that teach anything. The eight signals scored here live in the ledger at radar.ersiner.ai/signals, alongside the open calls and their resolution dates. First Scorecard: October 2026.

The Agents & Margins briefing is launching — until then, follow on LinkedIn

Sources

1. 1.The open and adjacent entrants of July and August 2026 — Y Combinator, QM (quartermaster), released 31 July 2026 under an MIT licence, repository read 18 August 2026 (source of "A multiplayer agent harness for work. In Slack and on the web", the per-person and per-room scoped memory, permissions, sandboxes and schedules, the accounting/legal/events/engineering deployment, and the authors' own "early and has bugs" warning); Block, Buzz, released 21 July 2026 under an Apache licence, repository read 18 August 2026 (source of "Agents are members, not bots", per-agent keypair and audit trail); and Nous Research, Hermes Agent under an MIT licence, with multi-agent rooms bundled on 18 August 2026 — rooms of agents, not of people. Y Combinator's launch post was read through search indexing rather than opened directly and is corroborated by the repository. marktechpost.com
2. 2.Anthropic — Introducing Claude Tag, 23 June 2026, and the product page read 18 August 2026. Source of "@Claude is multiplayer", the "hours or days" duration claim and the 65% internal-usage figure. claude.com/product/tag
3. 3.Anthropic — Claude Tag documentation and help centre, read 18 August 2026: overview (public beta, Team and Enterprise only, "Adding Claude to Slack doesn't add a per-seat charge", usage balance and spend limit), how it works (admin-provisioned service accounts, per-channel scope, memory scope, vendor-hosted sandbox) and the launch promotion ($25,000 per Enterprise organization, $2,500 per Team organization with 10+ paid seats, credits expiring 1 September 2026).
4. 4.Anthropic — plans and pricing (Team standard seat $20 billed annually / $25 monthly, premium $100 / $125, Enterprise "$20/seat. Usage cost scales with model and task."), and platform pricing documentation (Opus 5 $5/$25 per million tokens, Sonnet 5 $2/$10 with the 1 September 2026 increase cancelled, ~30% higher tokenization from Claude 4.7 onward, $0.08 per session-hour, $10 per 1,000 searches, $0.05 per container-hour, 1.1× United States inference multiplier). Both read 18 August 2026.
5. 5.Vendor pricing pages read for the shared-surface census, 18 August 2026 — Microsoft 365 Copilot Business ($18.00 user/month paid yearly, promotional from $21.00; "use of agents published to Microsoft 365 Copilot is included in their license"), Asana (Starter $10.99 and Advanced $24.99 per user/month billed annually, monthly credit allowances, no published price for AI Teammates) and Google Workspace (Gemini bundled across tiers; "AI Expanded Access" add-on listed with no price; US-dollar list prices not served to this workstation).
6. 6.OpenAI — Introducing workspace agents in ChatGPT, 22 April 2026 (research preview at launch, generally-available banner as read on 18 August 2026; free until 6 May 2026, then credit-priced), Introducing group chats in ChatGPT, 13 November 2025 with a 20 November expansion, and the help-centre retirement notice recording the 9 July 2026 wind-down. Pages read in a browser; the domain returns HTTP 403 to automated fetching. help.openai.com
7. 7.Microsoft — Announcing the new Work IQ APIs, 2 June 2026, generally available 16 June 2026; Ignite 2025, 18 November 2025 (Teams Mode, agents in Teams channels over MCP, Agent 365 as "the control plane for agents"); Entra Agent ID documentation; Copilot in Teams chats (generally available February 2025) and the Facilitator agent in Teams chats (generally available April 2025) — both outside the window, carried as historical precedent; and the FY2026 Q4 earnings call, 29 July 2026 (30M+ paid seats, "beyond per-seat to per-seat-plus-consumption").
8. 8.Andreessen Horowitz — Big Ideas 2026: Part 1, 9 December 2025. Alex Immerman, "The collaboration layer becomes the moat", in the vertical-AI section; Sarah Wang on the system of record losing primacy in the same collection. Published forecasts by a firm invested in the categories they favour.
9. 9.Asana — AI Teammates announcement, 25 September 2025, and Agentic Work Management, 4 June 2026 (source of "multiplayer mode" and "multiplayer coordination"); Arnab Bose, chief product officer, interviewed by Matthew Finnegan, Computerworld, 1 April 2026, source of "multiplayer by design" and the work-graph concession.
10. 10.Beatrice Nolan — Fortune, 23 June 2026: Cat Wu, Anthropic's head of product for Claude Code and Cowork, "Claude Code, Cowork, and chat are very single-player".
11. 11.Anthropic, the record before Claude Tag — Claude Code in Slack documentation, read 18 August 2026 ("runs each session under an individual user's account"; product shipped in beta 8 December 2025); Projects, 25 June 2024 (shared project knowledge and conversation snapshots — outside the window, historical precedent); and Claude for Small Business, 13 May 2026 (15 workflows and 15 skills through Claude Cowork, connectors into third-party finance, customer and signature records; no price stated).
12. 12.Harrison Chase, LangChain — Introducing ambient agents, 14 January 2025. Outside the evidence window and cited as the origin of the term: an ambient agent listens to "an event stream and act[s] on it accordingly". The publisher sells the orchestration and agent-inbox tooling the category requires.
13. 13.Kezia Jungco — TechRepublic, 24 June 2026, reporting Reuters: Rob Seaman, general manager of Slack, on making artificial intelligence "multiplayer" and on the capabilities having existed before the form factor.
14. 14.Shared-surface releases — Notion, 3.6 external agents, 1 July 2026; Salesforce, Slackbot general availability, 13 January 2026; Google Workspace at Cloud Next '26, 22 April 2026; Linux Foundation, Agent2Agent past 150 organizations, April 2026 — a self-reported membership count, not an adoption measurement.
15. 15.ICONIQ Growth — State of AI: Bi-Annual Snapshot, published January 2026; proprietary surveys fielded April 2025 and December 2025, roughly 300 executives at software companies building AI products, N=297 for pricing, N=269 for margins, N=202 for cost composition. The publisher invests in the companies it surveys and states that responses include some of its portfolio.
16. 16.Kyle Poyar, Growth Unhinged — The 2026 State of B2B SaaS and AI Monetization Report, 13 May 2026; more than 230 self-selected business-software and AI companies, fielded April–May 2026 (hybrid pricing 37% against 25% twelve months earlier; AI credits active 29%, planned 33%).
17. 17.Grok Bot, xAI — early beta opened 11 August 2026; a standalone personal application with one persistent cloud computer per account, whose group chats coordinate that account's own agents. The vendor's own site returns HTTP 403 to automated fetching and was not opened, so this rests on launch-day trade coverage reproducing vendor copy: Unite.AI (Evan Mercer, 11 August 2026) and Reworked (Siobhan Fagan, 12 August 2026). No independent testing exists.
18. 18.Kiteworks — aggregation of three agent-governance surveys, 29 July 2026: 1Password (n=1,000, fielded late May to early June 2026), Okta Global CISO Insights (n=306) and Kiteworks' own (n=459). Two of the three field periods are undisclosed in the aggregation; all three publishers sell into the gap they measure.
19. 19.Axios, 17 August 2026, and CNN Business, 18 August 2026 — Google's winning bankruptcy-auction bid of approximately $10 million for Spirit Airlines' corporate corpus, including roughly 500 million Microsoft Teams messages; announced, pending court approval; price press-reported, not published-primary.
20. 20.Cohort and outcome-pricing disclosures — Salesforce Q1 FY2027, 27 May 2026 (Agentforce ARR $1.2B, +205% year over year, disclosed as a supplemental metric); Atlassian Q3 FY26 shareholder letter, 30 April 2026 (agent credit usage +20%+ month over month, seats expanding, no agent revenue line); monday.com Q2 2026, 10 August 2026 (AI products at 17% of net new ARR); and HubSpot outcome pricing, effective 14 April 2026 ($0.50 per resolved conversation, $1.00 per recommended lead).
21. 21.The in-harness package economy, read 18 August 2026 — published counts from the largest third-party directory of harness skill packages (800+ packages, 20 curated bundles, advertised free) and observed listing prices on general creator-commerce platforms (roughly $19 to $167). Individual sellers and directories are deliberately not named: the finding is the category's price structure, not an evaluation of anyone's package, and no platform publishes category-level churn, refund or retention data. Counts are self-published and unaudited; the price sample is small and non-random.
22. 22.Cloudflare, the company-workspace pair of 5 August 2026 — Cloudflare OS: an open platform for agents, apps, and work (source of "every agent and app starts with access to nothing", of apps shared as blueprints from which each recipient creates an independent copy, and of each app holding "its own SQLite database, separate from the Cloudflare OS runtime"; the post says the platform is open source and names no licence, no price and no seat cost anywhere) and How we're rethinking work at Cloudflare with Cloudflare OS, by the company's chief information officer (source of "over 4,000 apps and tools" created in thirty days and of a user's session being "scoped to their existing permission set"). Both counts are self-reported and unaudited, and Cloudflare sells the Workers, Access and gateway products the platform runs on. Both read 18 August 2026.
23. 23.SemiAnalysis (Daniel Nishball, Dylan Patel and colleagues) — AI Value Capture — The Shift To Model Labs, 1 May 2026. Partly paywalled and read at headline and summary level; a modelled attribution rather than a disclosure, published by a firm selling research into the buildout it measures.
24. 24.Prior-art roadmaps — Mike Droesch and Darsh Patel, Bessemer Venture Partners, Roadmap: AI systems of action, 28 May 2025 (outside the evidence window, cited as prior art); and Foundation Capital, The $4.6T Services-as-Software opportunity, 18 August 2026. Both are investment theses published by firms invested in the categories they describe.
25. 25.Thomas W. Malone, JoAnne Yates and Robert I. Benjamin — Electronic Markets and Electronic Hierarchies, Communications of the ACM 30(6), June 1987; Thomas W. Malone and Kevin Crowston — The Interdisciplinary Study of Coordination, ACM Computing Surveys 26(1), 1994; and the organization-capital literature surveyed by Baruch Lev. stern.nyu.edu