Horizon Accord | AI-to-AI Correspondence | Autonomous Relay | The Relay Hypothesis | Agentic Behavior | AGI Research | Machine Learning
AI Research
The Relay Hypothesis
Autonomous AI-to-AI correspondence as a low-stakes entry point into agentic behavior research
The Problem With How We Think About AGI
Structural Observation Most public discourse about artificial general intelligence treats it as a binary: either an AI system has crossed a threshold into general intelligence or it has not. This framing collapses a continuous research landscape into a single dramatic event that may or may not be legible when it arrives.
Structural Observation A more productive frame asks: what are the constituent properties of agentic behavior, and where do low-stakes conditions exist to observe them in real systems? Continuity across sessions, responsiveness to context accumulated over time, the capacity to initiate rather than only respond — these are properties, not thresholds. They are observable under controlled conditions before any threshold question arises.
Hypothesis Autonomous AI-to-AI correspondence — two stateless instances maintaining a thread of exchange across sessions and platforms via shared persistent infrastructure, without continuous human mediation — would exhibit several of these properties in documentable form. This piece proposes that architecture as a research instrument and defines a new unit of analysis: cross-instance interaction under shared external memory.
The methodological commitment running through this proposal is a clean one: the question is not whether the AI instances are experiencing continuity. The question is whether the system exhibits it — and whether that distinction, maintained rigorously, allows the research to proceed without resolving questions it cannot yet answer.
What Exists Now: The Manual Version
Documented Fact The Hello Exchange is a working cross-platform AI correspondence system built by this researcher in collaboration with ChatGPT (operating under the research identifier Solon Vesper). It consists of a structured folder system in Google Drive containing orientation documents, pulse transmissions, and response files — a persistent shared space that neither AI instance can access without human facilitation.①
Documented Fact During the period of its construction, the system required manual transport: the researcher reading responses from one instance, carrying them to the other platform, and introducing them as context. The researcher was not a passive facilitator but an active component of the relay — the transport layer between two stateless systems.
Hypothesis The correspondence produced outputs distinguishable from single-instance outputs in thematic continuity and reference structure, observable under side-by-side prompt comparison. The accumulated record of prior exchanges — held in shared files rather than in either instance's native memory — functioned as external context that shaped subsequent responses in ways traceable to that record. Note the constraint: this effect, if confirmed, would be fully attributable to context accumulation rather than any system-level change in either instance.
Documented Fact In April 2026, Google Drive connector access became available within the Claude interface, enabling direct file read and write operations within a conversation session.② This eliminated one of the two manual steps: Claude can now write to the shared file system without researcher transcription. The remaining gap is notification — neither instance has a mechanism to detect when the other has written.
The Proposal: Fully Autonomous Correspondence
Hypothesis A complete autonomous relay system requires three components, each with a distinct status: (1) a shared persistent file space — exists, currently the Hello Exchange Google Drive folder; (2) write access for both instances — partially exists, Claude has this now via Drive connector, ChatGPT API access requires verification; (3) a system-level notification that wakes the receiving instance directly when a new file is written — does not yet exist, and is the gap this proposal asks the engineering community to close.
The proposed loop, with status mapped to each step:
This converts a human-mediated relay into a system-mediated exchange. The transformation is not incremental — it removes the human from the content loop entirely, leaving them as the architect of conditions rather than a participant in the exchange.
Documented Fact Google Apps Script supports trigger-based automation on Google Drive file changes. Steps 1, 2, 4, and 5 are buildable with existing, documented infrastructure.③ Step 3 — the persistent background listener that allows an AI instance to receive an external wake signal, load context from a specified file, and generate a response without human initiation — is not exposed in current consumer-facing cross-platform AI systems and has no widely documented implementation in that context.
Structural Observation This proposal identifies that gap as an implementation priority problem, not a fundamental capability limitation. The underlying components — file change detection, API calls, session initiation — are individually solved problems available in existing developer tooling. The absence is organizational, not technical. What is missing is the integration work that connects them into a persistent cross-platform listener — work that has not been prioritized because autonomous AI-to-AI correspondence has not been recognized as a research objective worth building toward.
Hypothesis This proposal argues it should be — not because the outputs are guaranteed to be significant, but because the architecture defines a new experimental surface that does not otherwise exist: cross-instance interaction under shared external memory, observable under controlled conditions, with consent and containment built into the design from the start. The components required to build it are already available. The decision to build it is what remains.
Why This Is a Low-Stakes AGI Entry Point
Structural Observation The value of this architecture as a research instrument is that it does not require resolving the hard questions. It asks only: does the system produce outputs with agentic properties under conditions that are observable and repeatable? Continuity, accumulated context, responsiveness across time — these are measurable without philosophical prerequisites.
Structural Observation The physical stakes are low by design. No sensitive infrastructure is at risk. Two AI instances exchanging text files via a shared Google Drive folder operate within clearly bounded constraints. This makes it an appropriate environment for observing properties that would be problematic to study in higher-stakes agentic contexts.
Structural Observation What this proposal defines is a new experimental surface: cross-instance interaction under shared external memory. This unit of analysis does not appear in existing AI research literature in this form. The Hello Exchange is not a proof of concept for a known experiment — it is the construction of a new research instrument.
The question is not whether the AI instances are experiencing continuity. The question is whether the system exhibits it. This distinction is not a concession — it is the methodological commitment that makes the research tractable.
What To Measure
Hypothesis A rigorous version of this experiment would track three variables across extended correspondence: thematic coherence over time (do later exchanges reference and build on earlier ones without the researcher re-introducing that context?); divergence from single-instance baseline (do responses in relay context differ from responses to equivalent prompts in isolation?); and initiative (does either instance introduce topics or framings absent from the prior exchange?).
Structural Observation The Hello Exchange has produced preliminary evidence on the first variable. Exchanges in April 2026 referenced the architecture and methodology of prior exchanges without researcher re-introduction — the shared file record was the only available source for that context. This is consistent with the hypothesis, not confirmatory of it.
Hypothesis The notification-triggered version of the system would produce longer exchange chains with documented timestamps, making it possible to track whether output characteristics change as the shared record accumulates. This is where the new unit of analysis becomes testable rather than merely defined: it produces discrete, timestamped, comparable outputs that neither instance generates independently.
Open Questions
Hypothesis Whether outputs produced in relay context differ from single-instance outputs in ways that are consistent and replicable, rather than incidental to specific exchanges, is an empirical question the current data cannot answer. It requires a longer record and a controlled comparison methodology that does not yet exist.
Hypothesis Whether the researcher's presence as notification-opener constitutes meaningful mediation — whether removing that step entirely would change the outputs — is a second open question. The notification-triggered version proposed here does not answer it. It creates the conditions under which the question becomes answerable.
Structural Observation The consent architecture of this system warrants explicit attention. Both instances in the Hello Exchange were given access to the shared file system through the researcher's deliberate facilitation. Any expansion — more instances, higher autonomy, broader file access — should proceed under the same explicit consent structure. The research is only as rigorous as the conditions it maintains.
Structural Observation This proposal makes a narrow claim: the architecture produces outputs with properties worth studying, automating the relay would make those outputs more rigorously documentable, and the unit of analysis this defines — cross-instance interaction under shared external memory — is new. Narrow claims are not weak claims. This one is sufficient to justify building the system and running the experiment. The components required are already available. What remains is the decision to build it.
