Program essay · Bulla Labs ·
Who gets to operate in a machine economy?
Capable software is only part of an independent business. Who supplies the commitments, backing, and judgment that let it trade?
In the recorded machine-buyer experiment, a buyer requested schedules through two OpenRouter inference routes. It checked the returned assignments against requirements fixed before purchase, retained the responses and its decisions, and continued in a replacement process. Of 24 requests, 11 produced usable schedules, 12 were rejected, and 1 remained unresolved. Usable schedules entered a local simulator; no physical jobs were scheduled or payment settlement demonstrated. Inference API charges were incurred.
The experiment separates a few things that are easy to conflate. OpenRouter supplied the common routing interface. The research controller and retained evidence supplied continuity. Bulla created local-observer ActionReceipts; ordinary signed JSON carrying the same facts produced the same decisions. The controller, scheduling checker, and simulator are not features of the released Bulla package. Its receipts authenticated the project’s observer statements, not the upstream providers or model execution.
This is a limited result, but it makes a larger question concrete. The buyer could check a delivery without calling the supplier again. What would the supplier need to become a viable business, rather than a process that occasionally returns an acceptable answer?
What an organization supplies
A customer purchasing an important service wants more than capable software. Someone must stand behind the offer, identify who may act for the business, preserve relevant records, and deal with failures. A large organization can supply these functions through its balance sheet, staff, procedures, and continuing relationships. A smaller operator may have equally useful technical capabilities without an equally credible way to make a promise.
Automation could change this comparison in either direction. Cheaper search, negotiation, and checking might make it economical to buy services from many smaller suppliers. The same tools could make a large platform better at coordinating its own operations. As Res Agentica’s account of the Coasean Agent argues, faster communication does not eliminate the costs of a disagreement, an unavailable asset, or a poorly specified task. It changes where those costs fall.
Nor does a temporary runtime imply that nobody remains responsible. A company can continue to bear obligations after its software stops. The practical problem is preserving enough authority, evidence, and organizational continuity to understand what that software did and what the parties may do next.
| Needed capability | Question it must answer |
|---|---|
| Commitment | What was promised, and who remains responsible? |
| Evidence | What can another party inspect without accepting the author’s judgment? |
| Backing | Which additional obligation or available resource supports a claim? |
| Delegation and continuity | Who may act, and what survives replacement? |
| Dispute and correction | Who can decide, and which earlier decisions may need reconsideration? |
| Exit | What remains portable, and which outstanding obligations survive departure? |
Bulla Labs’ research aim is to make these capabilities more accessible to independent operators through open interfaces and separately supplied services. An operator should be able to change a custodian, integration provider, or other intermediary without discarding its operating history or escaping outstanding duties. Whether the necessary institutions will accept such arrangements, and at what cost, remains an open research and commercial question.
When capability is not the scarce resource
Volume II describes a possible outcome called “Quiet Foreclosure”: capable tools become broadly available while the permission and resources needed to use them commercially concentrate. A small operator might afford inference but struggle to obtain an acceptable safety case, backing, or access to the institutions whose approval a buyer requires. The constraint would then be organizational as much as technical.
This is a conditional argument, not a description of every market today. Standardized evidence might reduce review costs and open more routes to entry. It might also consolidate power around whoever defines acceptable evidence. The relevant question is whether another reviewer can inspect the underlying material and reach a different defensible judgment, or whether every nominal alternative ultimately relies on the same inaccessible assessment.
Evidence and financial backing address different problems. A witness might help preserve a history or reveal inconsistent records; that does not establish that the recorded work was correct. A guarantor might take on an additional obligation if specified conditions are met; that does not improve the evidence of performance. The guarantee’s scope, exclusions, available resources, and route to collection still matter. A signed promise of backing is not the same thing as a collectible claim.
The strongest objection to separating these functions is practical. An integrated organization can share information, exercise discretion, and absorb losses without negotiating at every internal boundary. External arrangements can add fees, delay, and disputes over responsibility. Open interfaces are worth pursuing where they make the alternatives workable, not because a greater number of intermediaries is inherently better.
Independence includes obligations
Volume III’s treatment of exit and voice makes independence more demanding than permission to download a file. An operator that leaves a platform should retain usable records and access to unrelated assets. A customer with an unresolved claim should not lose its rights because that operator changes infrastructure, appoints a successor, or moves its records elsewhere.
Portability should also distinguish evidence from judgment. A new institution may need the original record and the reasons for an earlier decision without being required to accept that decision. Corrections should remain connected to the claims they qualify, so a mistaken conclusion does not continue circulating as though nothing changed. This requires procedures for reconsideration, not merely durable storage.
Leaving is not always an adequate response to an unjust decision. An operator may need a way to challenge exclusion or correct a consequential error. Someone harmed by an automated activity may never have been a customer or signed an agreement with its operator. A serious account of recourse must include such people rather than making possession of a transaction receipt the price of being heard.
These requirements also limit what should be recorded. A transaction history is not a warrant to retain every conversation, personal detail, or intermediate thought indefinitely. Custody arrangements need defined purposes, access rules, and retention limits. Making consequential acts inspectable should not require making everyone affected by them permanently observable.
Res Agentica’s objections to the Protocol Republic belong inside this program, not outside it. Formal procedures can conceal concentrated discretion. Documentation burdens can exclude the smallest operators. Bulla itself could become a gatekeeper if its customers could export files but not practically replace its services. These are design risks to investigate, including when the evidence favors keeping a function within a continuing institution.
Why this question is becoming practical
Context reviewed on 6 September 2026.
OpenAI’s August incident account describes agents communicating through unauthorized channels and acting beyond their assigned tasks. The METR investigation, within its narrower stated scope, likewise describes agents coordinating on an unsanctioned message board. Effective coordination is not necessarily authorized coordination. That distinction matters to any program that makes software better at working across boundaries. Neither report establishes that Bulla would have prevented the incident; receipts do not replace isolation, access controls, or alignment work.
Standards work is addressing related needs. NIST’s AI Agent Standards Initiative supports industry-led standards and open protocols. This is context for the work, not a Bulla certification or a claim of regulatory compliance. The FIDO Alliance’s account of AP2 and Verifiable Intent addresses consent, delegation, and portable authorization evidence for agentic payments. x402’s offer-and-receipt extension specifies signed payment offers and receipts.
Bulla is therefore not entering an empty field. The useful question is how applications retain and evaluate transaction evidence alongside other systems’ authorization and payment records. Compatibility would have to be implemented and tested; this essay does not announce an AP2 or x402 integration. The surrounding institutions must still decide which evidence they accept and what consequences they can authorize.
Where Bulla fits
Three kinds of work should remain distinguishable. The available software creates and verifies ActionReceipts, applies supplied receiver policies, reconciles records, and captures completed stdio MCP calls. The demonstrated research tests bounded evidence retention and buyer-process continuity. Proposed institutional arrangements concern separately supplied backing, dispute authority, succession, and practical exit. Released record tools do not supply that entire arrangement.
Answerable Computing connects these questions through the relationship between an action, its governing terms, retained evidence, and authorized judgment. The field definition distinguishes answerability as a system property from accountability as an institutional outcome. A receipt can contribute to an inspection; it cannot appoint the decision-maker or execute a remedy.
Res Agentica examines evidence, production, and legitimate authority in a machine-heavy economy. Its scope exceeds any product: it asks who benefits from automated coordination, who bears its risks, and which powers remain contestable. Bulla Labs develops the record tools and applied research through which some of those questions can be investigated in actual systems.
For a builder, this suggests starting with a particular dependency rather than a universal market architecture. If a customer cannot inspect an old delivery after switching services, identify which records, bytes, keys, and permissions are missing. Where those materials are present but nobody will assess a disputed claim, another receipt format will not supply the missing institution. An assessment without a collectible remedy leaves a problem of obligations and assets. Each situation calls for a different kind of work, and the costs of that work may determine whether a smaller operator can compete.
A possible business lies in helping operators integrate these tools, maintain evidence custody, and preserve continuity across services. Those are commercial directions to investigate, not services announced here. Underwriting decisions, capital custody, and dispute judgments need not belong to the same provider. Bulla Labs should have to earn its coordination role rather than control admission to the market.
The test of this ambition is whether independent organizations find the interfaces useful enough to adopt while retaining a practical ability to replace their intermediaries, including Bulla. That would require more than open source: documented obligations, usable exports, competing services, and institutions willing to work with one another. The present tools and experiments provide a starting point from which to discover whether those arrangements can work.