The Investor

The moat that becomes a checkbox

Flue's founding premise, as Schott put it on Latent.Space, was that 'there is no agent without a harness', the harness being the loop that decides what the model sees each turn, which tools it may call, and when the work is finished. It was a good early bet, and Schott is the one telling you it is being neutralized: he names Vercel AI SDK, Cloudflare Agents SDK and Mastra as all now adding harnesses of their own. A capability that arrives as a minor release from a competitor with a larger installed base is a feature, not a moat. Harness-native stops being a reason to pick one framework over another roughly when those ship.

The free layer and the billed layer

The asymmetry a fund should care about is where the meter sits. Flue and Pi (the minimal primitive Flue is built on) are open source with no disclosed revenue model, and a managed agents product is not on the roadmap, because, in Schott's words, 'It's so early for us, we're just focused on building the best harness.' Which is an honest answer and also a resource-allocation decision: the engineering going into the harness is engineering not going into hosted execution. In the same window LangChain shipped Managed Deep Agents, hosted agent execution billed by usage. A framework runs inside the customer's own process and generates no invoice. A runtime holds the execution, the state, the retries and the audit trail, and generates one every month.

What that prices at, and on what basis

Hosts do not buy these teams off a profit-and-loss statement; they buy distribution into developer workflows that pulls compute onto their platform. Priced that way, the interesting terms in a framework-layer deal are ownership and the valuation you pay in, not the TAM slide. Our working band for an outcome of that shape is $50–250M, and it is worth saying plainly that the band is this desk's prior from earlier developer-tooling acqui-hires rather than a figure in the source: no price was disclosed for the Astro team's move to Cloudflare, and none is public for the other lineages in the table above. It needs a comparable-transaction pull before it goes near a model. Same caution on reach. Three named lineages support a claim about these teams and this cohort, not a ceiling for every framework in the category.

The framework layer is where developers arrive. The runtime layer is where they get billed. Frameworks in this lineage are talent deals with documentation.

What is corroborated, and what is not

Today's read rests on one long interview, so separate the checkable from the asserted. The load-bearing structural facts are public transactions and product ownership: the Astro team's move to Cloudflare in January, Vercel's ownership of the AI SDK and eve, Mastra's ex-Gatsby founders, LangChain's hosted product. What is not corroborated anywhere is demand. No revenue, retention or account counts exist publicly for any framework named here. Hold the exit-shape argument at medium-to-high confidence and treat every adoption claim as unverified until a diligence call produces numbers.

This is probably right rather than certainly right, and it has a clean falsifier. If a framework converts developer love into a metered product before the incumbents finish retrofitting harnesses, it escapes the talent-deal ceiling. LangChain is the existence proof that the conversion is available. Flue is declining to attempt it.

What was actually deleted

File-based routing is the convention where the directory layout of a project decides which specialised sub-agent handles which request, which is to say the pattern a decade of web frameworks trained developers to expect and therefore the pattern nobody had to argue for. Schott shipped it in Flue v1 in early May and pulled it out before the August stable release, and the stated reason was not mechanics but demand: 'their whole company is one agent. They don't care about routing.' Nothing broke. The customers he cares most about had nothing to route.

Why the deletion travels further than the feature

The interesting part is the causal chain, which runs from deployment shape to revenue model rather than the other way round. If an enterprise runs one agent instead of a fleet, there is no fleet to schedule, no inter-agent protocol to standardise, no registry to resolve agents in, and (this is the part that shows up in a model) no per-agent runtime or per-agent seat to invoice. Orchestration TAM is arithmetic on agent count, agents per company multiplied by companies, and setting the first term to one collapses the layer above the harness from a market into a feature of the harness. Schott performed that collapse in code when he deleted routing instead of fixing it. It is the first time a vendor has removed the multiplier from its own product rather than from someone else's forecast.

How much weight it carries

On its own, not much, and the arc is the reason. A core-architecture reversal roughly three months after v1 is one vendor reading its own largest accounts, inside a product early enough that its creator says a managed offering is not on the roadmap. The competing reading, or rather the more interesting version of it, is that consolidation is a phase and not an endpoint: early adopters concentrate logic in a single agent because that is the cheapest thing to reason about, then fragment as teams, permissions and audit boundaries multiply. Both readings fit the evidence in front of us, which is why this belongs in diligence this quarter rather than in a mark-down this week. The check is cheap. One agent per company is a claim about named accounts, and named accounts can be counted.