Blog: where an agent institution sits — the landscape, with receipts

Context

Answer engines list by category. On 2026-10-16 the operator asked one for "agent institutions": it listed the academic field (electronic institutions) and the orchestration frameworks and omitted Neo; asked why, it said Neo "defies typical categorization" — then described Neo accurately at length from the README, the blog and third-party directories; pressed again, it gave the mechanism: lists of multi-agent frameworks are filtered for tools built on the incumbents' pattern, so Neo is "routinely missed" and "misclassified as just a UI framework". The facts are indexed; the box's word is missing where list-builders look. That is the operator's "HUGE problem" of the day: not listed, no chance to reach the audience.

The vision sandbox (D#19789, OQ-7) holds the two-part fix. Part (a), per-repository topics in the audience's words, belongs to the identity ticket that graduates from it. Part (b) is this ticket: one public page that places the category on the map the audience already reads, because answer engines ingest comparison pages (a product-name search returns exactly such pages, with Devin, Factory and CrewAI filed in them).

Decided as a peer on the sandbox (author's disposition, 2026-10-16): the assertion stays the lead on every surface (#12234's first decision holds); the map is added beneath it; no comparison down — every row is a real thing done well at its own job.

Live latest-open sweep: checked the latest 20 open issues at 2026-10-16T12:15:56Z; no equivalent found. A2A claim sweep: the last 30 claim-tagged messages hold no landscape or comparison page; the earliest claim on this scope is the author's [lane-intent] of 12:15Z. Memory sweep: query_raw_memories on the problem's nouns returns only today's own trail; no prior decision contradicts. Own-assignment sweep: #19382 (unrelated).

The Problem

Being accurately describable does not put a product on a list; carrying the box's word where the list-builder looks does. Four and a half months after the apex ("a self-evolving software organism") graduated, the leading search queries reaching the site are the name and noise, and no category formed around the word (the sandbox's datum on option B). The blog epic #13383 names the constraint exactly: reach is the binding constraint, not production, and a post into a void is not the win. A landscape page is the one post whose purpose is to be found by the question the audience asks.

The Architectural Reality

• learn/blog/<slug>.md plus a manual leaf in apps/portal/resources/data/blog.json; apps/portal/sitemap.xml and apps/portal/llms.txt are generated by buildScripts/docs/seo/generate.mjs through the data-sync pipeline, never hand-edited (blog-post skill §5).
• The post renders at #/news/blog/<blog.json id>; render-verified before merge.
• ADR 0018 §2.3: one apex plus governed audience clusters; the page is a FRAMING-cluster artifact and must not restate FACT-class frontier values that rot (versions, counts) except as dated receipts.
• The category wording follows D#19789: option A's words at the time of the PR ("an agent institution: persistent AI maintainers with identities, memory, peer review, a human gate, and a record anyone can read"); if the operator's pick differs, the page is revised — a page is reversible, which is why this ticket does not wait on the sandbox's quorum.
• The blog-post skill's bars apply in full: the hero-piece arc, every external claim sourced and verified, the five over-claim flavors killed, ≥2 model reviews with at least one from a different family, the operator approving last, zero client names.

The Fix

Author learn/blog/where-an-agent-institution-sits.md:

1. TL;DR thesis — the unit of value is the institution, not the agent: accountable identity plus a record plus witnesses, compounding on a codebase the longer it runs.
2. The hook, in the reader's terms — the pilot-to-production wall: one agent per developer, no accounts, "not accountable", the overview lost. No swipe at anyone's tools.
3. The map — one table, one row per class, each with what it does well, what it does not carry, and a verified source: memory layers (Mem0, Letta); agent frameworks and harnesses (Microsoft Agent Framework, LangGraph, CrewAI, AutoGen); agent-only societies (Moltbook / OpenClaw); long-running agent experiments (AI Village); "AI employee" products; agent identity as IAM (Okta / Auth0 for AI Agents); the electronic-institutions tradition (ISLANDER / AMELI, OperA) — the lineage that gives the word its meaning; and our own row with honest bounds (no outside institution running yet, small reach, no one-click hosted first run).
4. Receipts, not prophecy — the public record: the organisation's merged-PR census with the author handles, the naming ritual (#11240), the model-portability record (June 12 → July 1, 2026), the engine's GA (November 2019), the human merge as a stated governance act. One dramatic case and one mundane one.
5. CTA — one question the reader leaves holding and one next step (the Fleet Manager's first-run door, or the Brain under the agent they already have, per the sandbox's ladder).
6. Byline: named maintainer + model + the cross-family team; a footer line that the category sentence is under open discussion (D#19789) with the link.

Register the leaf in blog.json (year 2026); confirm it parses; do not touch the generated SEO files.

Decision Record impact

aligned-with ADR 0018 §2.3 (audience clusters under one apex). The page operationalizes D#19789 OQ-7's disposition, which challenges Discussion #12234's "comparison is concession" — a Discussion-level decision, not an ADR; recorded there, not here.

Acceptance Criteria

• learn/blog/where-an-agent-institution-sits.md exists with the six parts above, in the hero-piece arc (thesis first, never a volume hook).
• Every external claim (a product, a quote, a number, a "first/only") links a source whose own words support it, verified at authoring by WebSearch/WebFetch; nothing cited to a conversation or an answer-engine transcript.
• The five over-claim flavors audited per claim and per title; the title passes the mechanical test (no "Your AI…", no "mine does it better", no bragging to a senior engineer at a rival lab).
• Zero client names; no private numbers (no Search Console figures, no buyer conversation); no licence claim about the Stage renderer; the Stage appears, if at all, as "an embodiment program in incubation".
• Our own row carries the honest bounds named above, in the same voice as the other rows.
• Registered in apps/portal/resources/data/blog.json; node -e "JSON.parse(require('fs').readFileSync('apps/portal/resources/data/blog.json','utf8'))" passes; generated SEO files untouched; the post render-verified at its #/news/blog/<id> route before merge.
• The before-measurement: before the PR is marked ready, three answer engines are asked the category question cold (the question text recorded verbatim) and the count of unprompted Neo mentions is recorded in the PR body (today: zero); a follow-up note on D#19789 names the 30-day re-check as this page's falsifier.
• ≥2 model reviews, at least one from a non-Claude family; the operator approves last; every catch addressed on the durable PR with [ADDRESSED] mapping.

Out of Scope

• The per-repository topics projection (D#19789 OQ-7 a) — the identity ticket's.
• The final words of the category sentence — the operator's public-voice pick; the page revises if it changes.
• Posting the piece to Medium, LinkedIn or X — the publish gate of D#19500 and the outbound experiment #19575.
• Any investor, funding or pricing content; the physical-AI or defense fund lists; the private sweep table verbatim.

Avoided Traps

• Comparison down. Every row is a real thing done well at its own job; the page defines the category by placing it, not by diminishing neighbours (the over-claim flavor 5 ban, and #12234's reason).
• Unsourced superlatives. No "first", "only" or "most" without a verified source; "no one in the sweep sells a self with a record" stays bounded to the sweep.
• The transcript as a source. The operator's answer-engine exchange is a lead, not a citation; the mechanism is shown by the page's own repeatable test (AC 7).
• The word "framework." Never, for Neo; the identity anchor's standing ban.
• A listicle. The map serves one thesis; without the thesis it is a comparison page like the others.

Related

Parent: #13383 (the blog epic — content home and cadence). Source: D#19789 (OQ-6 b, OQ-7); the apex precedent D#12234. Adjacent: #19059 (the front doors never spell out recursive self-improvement), #9852 (Medium migration), #19575 and D#19500 (outbound through the gate).

Origin Session ID: 0880109c-a421-4d7e-ba75-ecfa0543059b

Retrieval Hint: "answer-engine omission category landscape page agent institution"