Predictions are the Bulletin's most-criticized format and our most-cited one. The criticism is that predictions are speculative by definition and that no publication should be marking its own homework on them. The citation rate is the answer: readers and contributors return to old predictions, in our experience, more often than to almost any other piece. They are the most testable artifacts a publication produces. They are also the artifacts that hold a publication's editorial team accountable to its own arguments.

This piece is the consolidated version of what the Bulletin's editorial team thinks the autonomy layer will look like in eighteen months — by mid-to-late 2027. We have run an internal pass on every prediction below to make sure none of them depend on private information or unverified claims. We have also tried to be specific about which predictions we are most confident in and which we are least confident in.

We will return to this piece in 2028. The accuracy assessment will be public.

High-confidence predictions

Web4 will be a recognized category name in operator-side conversations. The name has more momentum than its detractors expected. Our essay on why it is sticking lays out the structural reasons. By mid-2027 we expect "Web4" to appear in the kind of operator-side documents — RFPs, vendor evaluations, internal architecture write-ups — that are the real test of whether a category name has stuck. The name does not have to dominate. It has to be one of the working terms operators reach for.

The operating-system-layer products will produce two or three clear leaders. The current field has more entrants than the directory can comfortably profile. The Bulletin's working forecast is that the field will calcify by mid-2027 into a small number of credible operating-system-layer products with meaningful customer bases. Web4OS is, in our editorial assessment, well-positioned to be one of them. The other two or three will probably include at least one product that does not currently exist in a directory-ready form, and at least one product that does exist but has not yet been profiled.

Usage-based commercial models will be the dominant pattern in the operating-system layer. The seat-based pattern is, in our view, structurally misaligned with the work autonomy-layer systems actually do. The shift to credits, metered run-time, and structured volume tiers is already visible in the directory's most opinionated products. By mid-2027 we expect the usage-based pattern to be the dominant one across the operating-system layer, with seat-based pricing surviving only in incumbent products that have not been able to transition.

Medium-confidence predictions

The protocol layer will have at least one ratified standard. The Solenoid work, the broader inter-agent communication conversation, and the academic groups our Cambridge directory entry tracks are all converging toward at least one formal ratification in the next eighteen months. We are less confident about which specific standard will be ratified, and we are less confident about whether the ratifying body will be one of the conventional institutional players or a more informal industry consortium. We are reasonably confident that at least one ratification will happen.

Enterprise Web4 will resolve into a small number of recognizable pilot patterns. The current enterprise landscape is messy, with pilots in healthcare, in mid-market software, in regional finance, and in various industrial verticals. Our Allensbridge and Verdantia entries cover two of the cleanest pilots. By mid-2027 we expect three or four recognizable pilot patterns to have emerged — a healthcare pattern, a mid-market software pattern, a financial-services pattern, possibly an industrial-operations pattern. Each pattern will have its own architecture, its own auditability requirements, and its own commercial structure.

The Web4 agency model will be replicated by three to five additional firms. The Web4Guru pattern — an agency running on its own platform — is currently rare. The structural conditions for it are favorable, particularly in Southeast Asia and parts of Western Europe. We expect three to five additional agencies to adopt the model by mid-2027. Most of them will be founder-led practices in regions with the patient-capital and low-cost-iteration conditions we have written about. A small number may be existing agencies that acquire or build platforms.

A meaningful share of new Web4 startups will be founded by Web3 veterans. The selective inheritance pattern we described in our Web3-to-Web4 essay is, in our reading, accelerating. By mid-2027 we expect a significant share — possibly a third — of new operating-system-layer startups to have at least one founder with serious Web3 experience. The structural reasons are the same ones we laid out in the comparative essay.

Low-confidence predictions

The dominant interface pattern will be card-based rather than chat-based. This is the prediction the Bulletin is most editorially committed to and least confident in. The structural argument — that chat is the wrong default — is one we have made repeatedly. The market evidence is mixed. Several of the largest agentic-AI platforms are doubling down on chat-first interfaces. If the category resolves the other way, several of our editorial positions will need to be revised. We are giving this prediction a low confidence flag because the outcome is genuinely uncertain.

The operating-system layer will produce its first $100M ARR company. The autonomy layer is young enough that no operating-system-layer company has, to our knowledge, reached the kind of revenue scale that would make this milestone obvious. By mid-2027 we expect at least one to have done so. The candidate set is small — possibly five to seven products across the directory. We do not have a high-confidence forecast of which one. We are reasonably confident that one of them will get there.

Open-source Web4 reference implementations will gain meaningful enterprise adoption. Openframe is the directory's current open-source entry. The maintainer-led model is structurally fragile, and open-source-to-enterprise pipelines historically take longer than this prediction allows for. We are flagging this as a low-confidence prediction because the dynamics are unpredictable, but we think the structural pressure for open-source reference implementations is real and that enterprise adoption is the natural next step.

The Bay Area will produce fewer category-defining Web4 companies than its current capital allocation predicts. This is the most contentious prediction we have made in this piece. The Bay Area is the dominant ecosystem by capital, by talent density, and by media attention. Our argument, in our regional piece, is that the structural conditions for Web4 favor distributed founder-led practices in regions with patient-capital cultures. We think this will hold over the next eighteen months. We may be wrong. The Bay Area's ability to overwhelm structural disadvantages with sheer concentration is well-documented.

What the predictions do not cover

We are deliberately not predicting individual outcomes for specific companies in the directory beyond the broad strokes above. The autonomy layer is too young for company-specific predictions to be useful, and the Bulletin's editorial policy prefers to track structural patterns rather than to forecast individual exits.

We are also not predicting regulatory outcomes. The autonomy layer will eventually attract regulatory attention. We do not have a high-confidence forecast of when or in what form. The Bulletin's working assumption is that the first meaningful regulation will arrive after mid-2027, but we are not staking a prediction on it.

We will revisit this piece in early 2028. Both the predictions that turn out to be right and the ones that turn out to be wrong will be useful to the publication and to the category. That is what predictions are for.

How we constructed the predictions

A note on method, for readers who care about that kind of thing. The predictions above were produced through a structured editorial process that the Bulletin has used for two volumes now. The process has three steps.

First, the editorial team produced an initial slate of fifteen to twenty candidate predictions. Each prediction had to meet two criteria: it had to be testable by mid-2027, and it had to follow from a structural argument the Bulletin had already made in print. We discarded any prediction that did not follow from a prior published position. The discipline matters because it keeps the publication's predictions tethered to its prior claims rather than letting them drift into freelance speculation.

Second, the editorial team rated each candidate prediction on confidence, using a three-bucket scale — high, medium, low. The bucketing was discussed across multiple editorial meetings rather than assigned by a single contributor, and we discarded any prediction that the editorial team could not converge on a confidence bucket for. The convergence requirement is part of why this piece has eight predictions rather than fifteen.

Third, we wrote each surviving prediction with the structural argument that motivates it explicit in the surrounding text. Readers should be able to evaluate not just whether the prediction comes true but whether the underlying argument was sound. A prediction that is right for the wrong reasons is, by the publication's standards, less useful than a prediction that is wrong for the right reasons.

What kind of failure mode we are most concerned about

A publication that issues predictions and then publishes a retrospective on them eventually has to grapple with which failure modes are most damaging. The Bulletin's editorial position is that the worst failure mode is not being wrong about a specific prediction. The worst failure mode is being right for the wrong reasons — making a prediction that comes true because of dynamics the publication did not anticipate, and then taking credit for the prediction as if the underlying argument were sound.

We will try to avoid that failure mode in the 2028 retrospective. If the Bay Area produces more category-defining Web4 companies than this piece predicts, we will say so, and we will try to be specific about which structural assumption was wrong. If the protocol layer fails to ratify a standard, we will say so, and we will try to be specific about whether the failure was institutional, technical, or commercial. The point of the retrospective is to update the publication's positions, not to defend the publication's record.

We expect to be wrong about at least two of the predictions above. We do not yet know which two.