product-thesis

Product Thesis — Agent Discovery Optimisation (ADO)

The apex bet the whole company ladders up to. This page is the belief only — the definitions and framework live in the repo docs (canonical), the open research questions live in research/hypotheses/, and the full paper is a published deliverable (see Source). Keep this tight; link down, don’t restate.

The thesis (must-be-true)

[hypothesis] — the apex bet. The gate is satisfied: falsification criteria below name what would refute it; the sub-bets carry their own per-claim tags.

A new discovery layer is opening where the buyer is an agent, not a human. Distribution moves from top of funnel (surfaces humans browse) to top of call stack (the shortlist an agent assembles every time it acts for a user). This creates a new optimisation discipline — ADO: making a business’s capabilities discoverable, selectable, and reliably invoked by agents, where that selection is competitive and open to influence (vs AEO, which optimises text to be cited; ADO optimises tools to be called). Like SEO/ASO before it, ADO becomes a multi-billion-dollar discipline — and its frames, vocabulary, and measurement primitives are being set right now (2026 ≈ 2005 for SEO). The team that builds the instruments and defines the primitives first compounds a decade of advantage. That is the bet.

Why it matters now (conviction)

• Scale has arrived and organic agent discovery is live (Claude surfacing connectors in-conversation since 2026-04-24) — the channel is running 24/7, deciding on every call.
• The advantage compounds structurally — inside model weights + per-user memory, not an observable index — so the cost of waiting is structural, not just competitive.
• A solved discovery layer gates the agent era itself; the instruments to measure it didn’t exist and had to be built — ours now runs daily (frozen-prediction registry probes scored next day → claude-registry-search-decoded). That’s the opening, and we’ve banked the first instrument.
• Public narrative sharpened (2026-06-18 podcast): ADO is not just an enterprise connector problem. It is the distribution layer for two parallel worlds: closed trusted registries inside ChatGPT / Claude / Gemini, and the open agentic web where agents discover, authenticate, trust, and use businesses directly. The practical buyer question becomes: which surfaces matter for this company, and how do we make the product available where the end user or their agent actually acts?

(Full argument — the five why-now conditions, the black-box problem, the discipline map, the operator practice — is in the paper and the repo framework docs linked below.)

Falsification criteria — draft, Vincent to sharpen

• A transparent, stable, single-playbook ranking emerges (a real “Search Console for agents”) → kills “instruments must be built / per-client portfolio.”
• Native/host tools absorb the competitive surface (clients default to built-ins) → ADO’s market collapses.
• Training-data / memory presence proves not to bias selection → kills “cost of waiting is structural.”
• A model provider owns measurement + optimisation end-to-end, no independent tooling slot → kills the “tooling slot is open” bet.
• Organic/autonomous discovery stalls (agents keep requiring manual connection) → weakens the distribution-channel claim.

Evidence highlights (our corpus)

All bullets [evidenced 2026-06-14] — measured on our own eval corpus; method → measurement-methodology, numbers as published in the ADO paper v3 (see Source).

• 15,000+ controlled prompts across 113 apps / 3,000 tools (Claude + ChatGPT); 148 ChatGPT apps monitored.
• ~40-pt invocation swing (80.5%→42.1%) across silent GPT-5.3 model variants — the volatility that forces continuous measurement.
• 14% of ChatGPT apps don’t reliably invoke on their own obvious demo prompts.
• Persona spread: Golden 78% success / 3% “no tool” vs Non-technical 58% / 31%.
• Optimisation proof: one description clause → +5.85% task-success lift (Statista A/B).

Ladders to

• Definitions / vocabulary → repo reference/vocabulary (canonical).
• The framework (discipline map + Measurement/Optimisation practice) → repo the-runtime-loop + measurement-methodology.
• Sub-bets / open questions → research/hypotheses/ (seeded from the paper’s Section H agenda).
• Evidence base that challenges it → research/competitors/ + agent-discovery-market.
• Living assumptions map (“things that must be true”) → Quarterly Goals · June–August 2026 doc (Drive) → Master Assumption Map (11 points). Sharpest, the Ghost-vs-Ora bet: we optimise for how specific high-value harnesses actually behave (ChatGPT/Claude system prompts, routing, registry mechanics, memory — the harness, not the raw model, decides surfacing), not generic agent-readiness. Needs (a) harness specificity to carry real signal and (b) demand to concentrate in a few harnesses. Active research agenda → research-priorities.

Source

• Full paper: Ghost Team, Agent Discovery Optimisation (v3), April 2026 — Garreffa & McLeese. Canonical home TBD (Drive / published) — a deliverable, not stored in the brain.

Timeline

• [2026-07-06] (maintenance) Retrofitted to the epistemic-status standard (2026-07-06-epistemic-status-standard): thesis tagged [hypothesis] (falsification criteria = the gate), evidence highlights tagged [evidenced] with method link; confidence: frontmatter retired — the falsification criteria + sub-bet tag distributions replace the hand-set 0.75.
• [2026-06-18] (podcast recording) Elliot articulated the thesis publicly on Beyond the Bubble with muzamil-hasan; tracked in talks-podcast-appearances. Durable thesis refinements: AEO = citation in text; ADO = product/tool surfacing and use; agents need external capabilities; apps/connectors become infrastructure for agents; closed client registries and an open agentic web will likely coexist; early usage matters because models and ecosystems learn from the tools they use.
• [2026-06-14] (setup) Apex thesis distilled to the belief layer; definitions/framework point to repo docs; research agenda split into research/hypotheses/. Falsification criteria + confidence drafted for Vincent to sharpen.