The Product Desk

The sequencing carries more information than the multiple. Airtable launched HyperAgent, its platform for building and deploying AI agents, around February 2026. Days before the Bending Spoons announcement, HyperAgent was carved into a standalone company, HyperAgent Inc., disclosed only through an SEC filing with no comment from either party. Founder-CEO Howie Liu plans to run it full-time once the deal closes, with rights to raise its own capital. Airtable investors get a stake.

Separate the thing being pitched from the thing being bought. An acquirer paying all cash for an installed base looked at a six-month-old agent platform and declined to take it. That is a cleaner verdict on bolt-on AI than any multiple.

Where the money actually came from

The recovery was balance-sheet driven, not business driven. Roughly $900M of unspent cash sits on Airtable's books, a byproduct of raising $1.3B over 14 years, distributable on top of the price. Bain Capital Ventures' Aaref Hilaly estimates about $850M reaches common shareholders, of which roughly $127M is left for about 900 employees after a ~20% founder stake. That is near $141K a head against expectations set by an $11B paper mark. Hilaly's framing, per The Information's Dealmaker reporting: common holders get "something from this, just nowhere near what they were hoping for."

The efficiency numbers make it worse, not better. At north of $430M of implied annualized revenue across ~900 people, that is roughly $480K of revenue per head, genuinely good by SaaS standards. Efficiency did not set the price. Growth rate and a credible AI attach story did.

Where the sources disagree, and why it decides the move

Both readings support the same first action, which is why the disagreement does not need resolving before anyone moves. Enterprise buyers reopen settled tooling decisions during ownership transitions and almost never otherwise. The window closes the moment Bending Spoons names product leadership and publishes a roadmap.

The uncomfortable part for the backlog

Airtable ranked No. 2 on The Information's enterprise software acquisition-target list, which makes this comp the front of a queue and an anchor that drags every private SaaS negotiation downward. AI-application capital, meanwhile, is untouched. A customer feedback startup is in talks for a $125M round led by an Anthropic investor. Cheap distribution is buyable. Slow growth with an unmetered AI narrative is fatal.

The forcing function that follows is uncomfortable but cheap: for every AI initiative on the roadmap, name the retention, expansion, or win-rate number it moves and the date it must move it. Usage volume does not count as an answer. Airtable shipped an agent platform and then sold under 3x revenue with the agent platform removed from the transaction. If your AI work cannot name its metric, Finance will find this comp before you do and cut the line without product input.

An acquirer paid all cash for the installed base and left the AI platform on the table. That is what a non-accretive AI narrative looks like when someone finally prices it.

A non-developer asks for a deck and gets a deck. She never sees a commit or a diff, and that is the design. The mechanics matter more than the merge date. Inside ChatGPT Work, a task gets an isolated microVM with a working directory at /workspace/scratch, installs dependencies, keeps databases, grinds for hours, and emits artifacts: docs, sheets, slides, and Sites, hosted web apps it can share by URL and keep updating. Pro accounts get 8 CPUs, 20GB RAM and 64GB disk. Plus gets 14GB. Work runs on the Codex harness with git controls and diff traces stripped out, so a non-developer never realizes they are driving a coding agent.

One caveat, stated plainly: this is a single outsider reconstruction, published in Latent.Space by Shlok Khemani, who says the architecture "could look very different a few weeks from now." High confidence on the direction. Perishable on the internals.

The discovery failure is the whole business case

Here is what Work actually does. It routes flawlessly to plugins a user has already installed, and it never suggests one that is missing. Asked to find flights and hotels, it chose web search over the available travel plugins. Naming Expedia outright did not surface the plugin either, despite the plugin likely using fewer tokens, returning better results, and enabling direct booking. Here is what plugin teams tell themselves instead: the directory holds 1,000+ plugins, so the directory is distribution. The path into it is user awareness, and user awareness belongs to the plugin builder.

So a roadmap line reading "ship MCP server, get distribution from OpenAI's directory" has no business case today. Re-scope that work as retention depth for users already acquired, and drive installs from owned onboarding and lifecycle surfaces. The bet worth holding: with per-user microVMs and hosted browsers to underwrite, OpenAI has enormous incentive to fix commerce routing, and the plugins with real transactional depth collect the demand when it does.

Two adjacent constraints that hit owned surfaces first

• Datacenter origin is a hard capability ceiling. Work's Chrome runs as a separately hosted service. Amazon US rejected it as an unsupported "session or client" and Google Photos timed out repeatedly. Both tasks succeeded in local mode. Every revenue surface a team owns, checkout, account, media export, now needs a written decision: block, throttle, or sanctioned path.
• The canonical-source bug is a pattern, not a one-off. An uploaded file exists twice: a working copy in the thread and a canonical item in the Library. They do not synchronize. If one thread edits the Library version, another silently reads its stale copy on resume. A frontier lab shipped that into a product heading for a billion weekly users. Assume any in-house agent file model hits the same wall.

The channel nobody instruments

ben's bites points at the other half of the same problem. Agent recommendation is now a channel. Supabase Evals made agent-buildability a public benchmark, scoring how well Claude Code, Codex and OpenCode build with Supabase against real tasks, and the announcement pulled 327K views. Both sources land on one mechanism. The moment of choice moved inside an agent, and no dashboard reports the loss when an SDK confuses one.

The primitive shift is the part that belongs in a PRD. If the unit of interaction does not persist state, produce a durable artifact, and survive a device switch, the north-star metric is measuring activity rather than completion. Messages-per-session says nothing about whether the user got what they came for.

A directory of a thousand plugins with no path into it is not a channel. The install funnel is yours, and so is the cost of pretending otherwise.

A security architect sat through an agent demo this quarter and never asked about model quality. She asked who could revoke the agent, and how fast. That instinct is what Palantir's quarter is measuring, and the composition matters more than the headline. U.S. government sales grew 90% to $809M and total revenue grew 93% to more than $1.9B, but bookings outran both: $2.1B in U.S. contracts closed in a single quarter, up 153%, with full-year guidance raised to $8.15–8.158B from roughly $7.66B in May, a second consecutive raise. Bookings accelerating faster than revenue means contracts are being signed ahead of consumption. Budget is being committed now.

Separate the thing being pitched from the thing being bought. Per The Information's read of Karp's shareholder letter, what closed those deals was not capability. It was insulation: the argument that enterprises are "awakening to the risks of handing the creators of the language models the keys to their institutions." Anthropic's history of shipping products that compete with partners like Cursor and Figma gives that objection teeth. The Information also flags the obvious: it is an all-too-convenient message from a company whose stock has fallen this year.

The gap this opens is visible in the same week's launches

Write access is the capability inflection, not reasoning, and only one of those four shipped the governance to hold it. SentinelOne's underlying system already handles more than 8,500 critical investigations per day, with GA expected later this quarter, so the thing it is selling is reversibility rather than autonomy. Asana reached the same position from the productivity side: agent memories stay bound to the permissions of their source data, with model routing by task complexity.

What goes in the PRD template this sprint

Four fields, cheap to specify now and expensive to retrofit: admin-configurable action limits, per-action reversibility or a documented compensating action, an immutable decision trace including verdict rationale, and an approval-gate policy by action class. Per-agent identity sits on top of all four. Block just turned "one shared vendor-issued API key for every bot" into a named architectural liability, which is the kind of language that reaches an RFP before a standard stabilizes.

The counter-evidence keeps this honest, and it comes from usage rather than a pipeline slide. Coinbase built its own agent and Claude Code is still the most-used tool among its 2,500 engineers using AI coding tools, with senior director Chintan Turakhia saying there is "no crisp answer" on when engineers pick one over the other. Incumbents lose pricing power long before they lose usage, so a governance wedge wins procurement without winning daily habit. Two axes decide where a roadmap sits: whether the agent acts unattended, and whether each action is reversible with a trace to prove it. Price and position accordingly, because the demo moment that closes an enterprise review is revoking a compromised agent's access live, on screen, in under a minute.

Buyers stopped asking whether the agent can act. They are asking what happens when it is wrong, and who can prove it.