← Back to the programme

Population-Scale Agency and the Institutions of the Synthetic Economy

A five-track research agenda. This is a statement of intended work: no track has been completed, and several depend on access the programme does not have.

Read the research note Case SE-001

Document: Research agenda
Version: v0.1 — 12 October 2026
Tracks: Five; none completed
Priority: Track 4 — evidence and contestability

Standing No track is complete. This is intended work.

Standing of this document

This is a statement of intended work. No track below has been completed, and several depend on access to data the programme does not have and has no present route to obtaining. Where that is so, the track says it under Methodological obstacles. Nothing in this agenda should be read as a finding, and no hypothesis stated here has been tested.

The agenda follows from Research Note No. 1 (RN-001, 12 October 2026) and from Working Paper No. 1 v0.3-RC1, which remains a private review candidate. Where a track would bear on a WP-001 proposition, that is noted; such bearings are open issues for a future controlled revision of WP-001 and are not applied to it.

Tracks are connected but not sequential. Track 4 is the programme's current priority because RN-001's weakest evidence is there and because its first study is the one most nearly within reach.

Track 1 Unit of control against unit of attribution.

Track 1 — Delegation and authority

Research problem. Existing principal–agent theory and agency doctrine take the relationship as the unit of analysis. In documented enterprise agent architectures, authority is set at the level of a template — an identity blueprint from which many agents inherit scope — while action remains attributed to the individual agent. The problem is to determine whether the divergence between the unit of control and the unit of attribution has analytical consequences, or whether it is an implementation detail that existing frameworks absorb.

Primary questions.

When authority is inherited from a template and evaluated at authorisation time, what is the correct unit of analysis for economic and legal purposes?
Does dynamic, machine-evaluated scope differ in any legally material way from static delegated authority, or only in tempo and inspectability?
Under what conditions does an agent population become orphaned — holding live permissions with no accountable human — and what is the incidence of that condition in practice?
What is the operative difference between an agent acting on behalf of a user and an agent acting autonomously, where both are provided for as distinct authorisation policies?
Is there a defensible threshold at which a delegation architecture should be treated as an organisational structure rather than as a set of relationships?

Existing theoretical foundations. Principal–agent theory; Restatement (Third) of Agency on actual and apparent authority and sub-agency; capability-security and attenuable-token literature; organisational economics on delegation and authority within the firm; the UNCITRAL Model Law on Automated Contracting attribution rule.

Candidate evidence. Vendor documentation for agent identity platforms across at least three providers; published reference architectures; enterprise policy documents where obtainable; incident write-ups involving misconfigured permissions at population scale.

Testable hypotheses.

H1.1 Blueprint inheritance depth correlates with the time required to answer "was this action within scope?" for a randomly selected historical action.
H1.2 A material fraction of agent identities in mature deployments lack an assigned accountable human, notwithstanding that assignment is described as important.
H1.3 Organisations converge on a small number of blueprints relative to agent count, so that effective policy diversity is far lower than agent diversity.

Methodological obstacles. Enterprise permission configurations are confidential and security-sensitive. Vendors publish architecture, not deployment statistics. H1.2 and H1.3 require data the programme cannot presently obtain; they are stated so that someone with access can test them.

Potential empirical studies. Comparative documentary analysis of three or more agent identity platforms against a fixed schema of conferral and provenance features — feasible now, requiring no access. A practitioner survey on orphaned agents and blueprint counts — feasible with a professional body as partner. Instrumented synthetic deployment measuring scope-resolution time against inheritance depth — feasible in a laboratory.

Next outputs. (1) A comparative conferral/provenance feature matrix across platforms, which is also the replication RN-001's critique C8 demands. (2) A short note defining the unit-of-control / unit-of-attribution distinction formally enough to be argued with.

Track 2 Configuration cascade against emergent coordination.

Track 2 — Economic coordination and systemic behaviour

Research problem. The programme's position on emergent coordination is deliberately conservative: demonstrated in controlled settings, contested external validity, with transfer to deployed markets the exposed step. RN-001 adds a mundane mechanism that requires no emergence — correlated behaviour by shared configuration. The problem is to determine which mechanism accounts for aggregate effects that actually occur, and to resist the programme's own preference for the more interesting answer.

Primary questions.

Can population-wide correlated action arising from shared configuration be distinguished empirically, after the fact, from correlated action arising from learning or emergence?
What aggregate harms are reachable by configuration cascade alone?
Does the existing experimental literature on algorithmic coordination transfer to intra-firm agent populations, which differ from the inter-firm setting it studies?
What early indicators, if any, distinguish a developing cascade from normal variance?
Where an aggregate harm has no single responsible dyad, what forum is in fact available?

Existing theoretical foundations. Calvano et al. (2020; 2023) on algorithmic pricing and the genuine/spurious collusion distinction; multi-agent reinforcement learning on social dilemmas; agent-based computational economics; systemic-risk and cascade literature from financial stability; normal accident and high-reliability organisation theory; change- management and configuration-management practice.

Candidate evidence. Published post-incident analyses of automation-driven outages and trading disruptions; regulatory incident reporting where available; the appellate record on algorithmic coordination as already recorded in the Ledger; controlled experiments.

Testable hypotheses.

H2.1 Configuration-driven correlation and learned correlation leave distinguishable signatures in the timing distribution of agent actions — the former near-simultaneous and tied to a change event, the latter diffuse.
H2.2 In intra-firm populations, the dominant source of aggregate anomaly over any given period is configuration change rather than learned behaviour.
H2.3 Existing operational controls — change windows, staged rollout, circuit breakers — substantially bound configuration-cascade harm, so that the residual problem is narrower than the programme has assumed.

Methodological obstacles. Incidents are under-reported and legally sensitive. Distinguishing mechanisms after the fact requires exactly the provenance records whose absence Track 4 studies — the two tracks are therefore coupled, and Track 2 may be blocked by Track 4's findings. Laboratory populations may not reproduce the relevant dynamics.

Potential empirical studies. A signature-detection study on synthetic populations under both mechanisms. A structured review of published automation-incident analyses coded for mechanism. A desk study of whether standard operational controls would have bounded each reviewed incident.

Next outputs. (1) A note stating the configuration-cascade mechanism precisely and setting out its detection signature — explicitly offered as a competitor to the programme's own emergence emphasis. (2) A coded review of published incidents.

Track 3 Four states of recognition; revocation and in-flight commitments.

Track 3 — Legal recognition and liability

Research problem. Four states — technical authorisation, attributable authority, contractual recognition and sovereign legal recognition — are routinely conflated, and they come apart in both directions. Two of them scale automatically and two are resolved by human institutions one dispute at a time. The problem is to establish what follows from that difference in scaling behaviour, and to determine where existing doctrine genuinely runs out rather than merely becoming inconvenient.

Primary questions.

What is the legal effect of a mandate that executed technically but exceeded the scope the principal intended, where the counterparty relied on it?
What becomes of commitments already initiated by downstream agents when upstream authority lapses or is revoked?
How do intermediaries — platform operators evaluating authorisation, scheme operators defining mandate formats — sit in the liability structure?
Does the attribution rule for automated systems extend cleanly to multi-hop machine delegation, or does it assume a single system acting for a single person?
What would sovereign recognition of a foreign-constituted machine mandate actually require as an instrument, and which existing instrument comes closest?

Existing theoretical foundations. UNCITRAL Model Law on Automated Contracting and its attribution rule; Restatement (Third) of Agency; private international law on recognition; intermediary liability doctrine; scheme-rule and payment-network contractual allocation; Kolt on governing AI agents.

Candidate evidence. The Model Law and its travaux; UNCITRAL Working Group IV session documents; payment scheme rules where public; decided cases involving automated authorisation disputes; comparative e-signature and digital power-of-attorney instruments as the nearest recognition analogues.

Testable hypotheses.

H3.1 Existing doctrine resolves the great majority of single-hop disputes without difficulty, and the residue concentrates in multi-hop and cross-border cases.
H3.2 Contractual recognition through scheme rules does more practical work than any public-law instrument, and will continue to.
H3.3 No current instrument track is working on sovereign recognition of foreign-constituted machine mandates — confirmed for one track, untested for others.

Methodological obstacles. Doctrinal analysis across jurisdictions is slow and requires qualified lawyers the programme does not employ. Scheme rules are frequently confidential. Disputes settle, so the case record systematically under-represents the hard questions.

Potential empirical studies. A doctrinal study of in-flight commitments on revocation, in two or three jurisdictions, conducted with commercial lawyers. A survey of publicly available scheme rules for agent-mandate provisions. A mapping of instrument tracks beyond UNCITRAL — Hague Conference, regional bodies, sectoral regulators — to test H3.3.

Next outputs. (1) The revocation and in-flight-commitment study, which is the sharpest unanswered legal question RN-001 raised. (2) An instrument-track map testing H3.3.

Bearing on WP-001. This track bears directly on WP-001's Proposition 2 and on its stated falsifier. Findings are recorded as open issues for a controlled revision; WP-001 is not amended from this agenda.

Track 4 The programme’s current priority. Mandate provenance.

Track 4 — Evidence, verification and contestability

Research problem. To contest an autonomous action a party must establish the action, the acting identity, the mandate, the scope at execution time and the approval. RN-001 finds the inputs to all five specified in platform documentation, and the retained linked record — mandate provenance — not specified there, with the interoperable form of it unadopted in the standards layer. The problem is to determine whether this is a real evidentiary gap or an artefact of documentation, and if real, what it costs.

This is the programme's current priority track. RN-001's weakest claim lives here, and its first study requires no privileged access.

Primary questions.

Across platforms, is mandate provenance specified, and at what granularity?
What does it cost — in time and in records consulted — to reconstruct authorisation for an arbitrary historical action, as a function of population size and delegation depth?
Does the declared/verified evidence distinction drawn in the standards work correspond to any difference a tribunal would recognise?
Who chooses the evidence model in practice, and is it the party who would be the defendant?
How do existing evidence doctrines — business records, spoliation, adverse inference — allocate the risk of unreconstructable authorisation, and is that allocation sufficient?

Existing theoretical foundations. Evidence law on business records, authentication of system-generated records, spoliation and adverse inference; audit and assurance standards for system records; provenance research in data management; RFC 8693 and the delegation-path work built on it; digital-forensics reconstruction methods.

Candidate evidence. Platform documentation across providers; the actor-chain Internet-Draft and any successor or adopting work; published audit standards; decided cases on admissibility and sufficiency of system-generated authorisation records; instrumented laboratory populations.

Testable hypotheses.

H4.1 Reconstruction cost rises more than linearly with delegation depth and approximately linearly with population size, under current record designs.
H4.2 Conferral features are specified at finer granularity than provenance features across platforms — the generalisation of RN-1 beyond its single-vendor base.
H4.3 Where provenance is configurable, deployments default to the weaker evidence model.
H4.4 Existing evidence doctrine resolves the allocation question adequately for single-hop domestic disputes and inadequately for multi-hop cross-border ones.

Methodological obstacles. H4.3 requires deployment configuration data, which is confidential and security-sensitive; it may be approachable only through practitioner survey with strong anonymity. Laboratory reconstruction cost may not transfer to production environments with mature tooling. The legal half of this track needs an evidence lawyer.

Potential empirical studies. (a) The comparative documentation study testing H4.2 — feasible immediately, no access required, and the programme's intended next piece of work. (b) An instrumented reconstruction experiment on a synthetic population testing H4.1. (c) A doctrinal review of admissibility and sufficiency for system-generated authorisation records.

Next outputs. (1) The comparative conferral/provenance study, which simultaneously replicates or refutes RN-1. (2) A working definition of mandate provenance precise enough to be assessed against, and to be wrong about.

Track 5 Consolidation, switching cost, standards governance.

Track 5 — Permission infrastructures and institutional power

Research problem. Permission to transact is becoming a layer with identifiable controllers. Two consolidations are documented: commercially, within a vendor's identity estate under a converging registry and a per-user licence; institutionally, within a standards body that has received the major agent-payment contributions and published no specification. Consolidation is the precondition for rents, not evidence of them. The problem is to determine where economic power in this layer actually sits, and to avoid measuring the wrong variable.

Primary questions.

What does it cost to move an agent population from one recognising authority to another, and how does that cost scale with population and inheritance depth?
Is the economically significant variable the price of permission or the cost of switching the authority that recognises you?
What governance does a converging standards body have, and who is represented in it?
Under what conditions does a recognition layer become a bottleneck capable of extracting rent, as distinct from a commodity?
Does consolidation of the permission layer raise or lower the cost of contestability?

Existing theoretical foundations. Coase on rights as factors; Barzel on delineation costs; Rochet–Tirole on two-sided access pricing; Laffont–Tirole and Stigler on capture; switching-cost and lock-in literature; standards-setting governance and essential-facility analysis.

Candidate evidence. Published licensing terms; standards-body governance documents and membership; migration case studies where published; the documented registry convergence and licensing structure; scheme participation rules.

Testable hypotheses.

H5.1 Switching cost rises with blueprint inheritance depth, because inherited scope must be re-expressed in a different vendor's constructs rather than merely re-granted.
H5.2 Rent in this layer accrues to lock-in rather than to posted price, and is therefore invisible in any price series.
H5.3 Standards-body governance concentrates in the contributing firms, measurable through working-group chair and editor composition over time.

Methodological obstacles. Licensing terms are negotiated and confidential. Migration costs are rarely published and never independently verified. H5.2, if correct, implies the programme's own proposed authority-price index would measure the wrong thing — a methodological problem the programme must resolve before investing in that instrument.

Potential empirical studies. A structured migration-cost study with practitioners who have moved agent estates. A longitudinal governance study of the relevant working groups, coding chairs, editors and contributors — feasible now from public records. A desk analysis of published licensing structures.

Next outputs. (1) The working-group governance baseline, taken now so that later concentration is measurable against it. (2) A note on whether an authority-price index or a switching-cost index is the right instrument — explicitly a challenge to the programme's own prior methodological commitment.

Bearing on WP-001. This track bears on WP-001's recognition-cost and permission-rent programme. Recorded as an open issue for controlled revision; not applied to WP-001 here.

Dependencies Which tracks block which.

Cross-track dependencies

Track 2 depends on Track 4: distinguishing correlation mechanisms after the fact requires the provenance records whose availability Track 4 is studying.
Track 3 and Track 4 are coupled on the legal side: the sufficiency of evidence doctrine is simultaneously Track 4's question 5 and Track 3's test of where doctrine runs out.
Track 5's switching-cost hypothesis depends on Track 1's finding about inheritance depth.
Every track depends on the comparative platform study in Track 4, which is why it is first.

Sequencing What is feasible now, with partners, and not at all.

Sequencing

Immediate, no access required: the Track 4 comparative conferral/provenance study; the Track 5 standards-governance baseline; the Track 3 instrument-track map.

Requires partners: practitioner surveys on orphaned agents, default evidence models and migration costs; doctrinal studies requiring qualified lawyers.

Requires access the programme does not have: anything resting on deployment configuration or incident data. These items are stated so that they can be taken up by researchers who do have access, with attribution to the originating framework.

Conceptual framework developed by Steven Alber. Research assistance is AI-assisted and disclosed in each document. This agenda has not been peer-reviewed and describes intended work, not completed work.

Correspondence and criticism: research@thesyntheticeconomy.org