Research & valuation · September 2026
The agent economy in 2026 has a peculiar shape: capital has never been more concentrated, adoption has never been wider, and almost nobody can say afterwards what their agent actually did. This page researches the markets the [&] portfolio sits inside, prices its domains with a method you can disagree with line by line, and says which of its bets the market is about to reward, which it is about to commoditise, and what should be sold first.
01 · The thesis in one diagram
The five markets below are usually analysed separately. Put them side by side and they point at the same missing thing: a record of what an agent was allowed to do, what it did, and whether the two agree, that a second machine can check.
Gartner's 40% cancellation forecast, a 17% replay rate among organisations running agents, and prompt injection in a third of sampled skills are the same finding three times: the agent stack can produce actions faster than it can account for them. The [&] portfolio's product is the accounting — composition that refuses instead of degrading, receipts a stranger can re-derive, and a world that is a graph rather than a snapshot.
That is a narrow claim and it is meant to be. The portfolio does not compete on model quality, on hosting, or on retrieval accuracy, and this page will show that it would lose on all three. It competes on the part of the stack that the compliance calendar, the security lists and the cancellation statistics are all now pointing at, and it competes there with running code whose claims were derived rather than typed.
02 · Five markets, sized by named sources
Analyst forecasts for “agentic AI” vary by a factor of ten depending on what is counted. Ranges are shown rather than a point, and the definition each firm uses is named where it changes the answer.
MarketsandMarkets' AI-agents series is the most-quoted band. Its narrower “enterprise agentic AI” cut runs $5.3B to $24.5B over the same years. Grand View Research's wider definition reaches $183B by 2033. Gartner expects 40% of enterprise applications to embed task-specific agents by end of 2026, from under 5% in 2025.
Mem0 $24M (Series A, Oct 2025), Letta $10M, Cognee $7.5M, Hindsight $3.6M, Zep $3.3M. Retrieval quality has converged: Mem0 94.4%, Mastra 94.87%, OMEGA 95.4% on LongMemEval with cloud readers. The category's next fight is provenance and cost per recall, not accuracy.
The EU AI Act now requires automatic, lifetime event logging for high-risk AI, retained at least six months on both provider and deployer sides, with penalties to 3% of turnover. OWASP's Agentic Top 10 (Dec 2025) names memory poisoning, tool misuse and rogue agents. A Q1 2026 survey of 420 organisations found only 17% could reconstruct an agent's tool-call sequence after the fact.
Trade coverage puts enterprise inference at the edge or on-premises well above its 2023 level, and “sovereign AI” has moved from a compliance niche to national infrastructure budgets. Small models now do real reasoning in a few billion parameters on hardware people already own. The Cloudflare Agents SDK, by contrast, makes a hosted stateful agent cost nothing when idle.
AMI Labs (~$1B), World Labs ($1B), General Intuition ($320M), Odyssey ($310M), Decart ($300M). Their “world” is a learned simulator of physics and causality that agents train inside. It is not this portfolio's “world”, which is a machine, its state and the record of how it got that way, in three named sizes.
Up from 18% in January; 47% in the United States. Codex grew about fivefold to 16%. Roughly a third of Claude Code's heaviest users generate more than 80% of their code with agents. Every one of those sessions leaves a transcript on disk.
03 · The standards layer settled this year
In December 2025 Anthropic, Block and OpenAI formed the Agentic AI Foundation under the Linux Foundation and donated MCP, goose and AGENTS.md to it, with Google, Microsoft, AWS, Cloudflare and Bloomberg backing. On 17 August 2026 Google's A2A joined the same foundation. Agent-to-tool and agent-to-agent are now governed in one place, with an enterprise roadmap that prioritises audit trails, SSO and retry semantics.
MCP's own numbers: more than 10,000 published servers and around 97 million monthly SDK downloads. The skills layer above it grew from one registry to eight marketplaces in six months and now indexes somewhere between 800,000 and 1.9 million SKILL.md files depending on who is counting. Quality is the story there: an audit of 22,511 skills found 140,963 issues, and Snyk found prompt injection in 36% of a sample.
The [&] Protocol's position is that it sits above both: it composes capabilities and compiles the result to MCP configurations and A2A Agent Cards. That claim got cheaper to make and easier to test this year, because the targets stopped moving. It also got more necessary, because the thing the foundation is standardising is transport, and the thing the skills audits are finding is that nothing at the transport layer refuses a bad part.
| Layer | Standard | Portfolio artifact | Rung |
|---|---|---|---|
| Agent ↔ tool | MCP (AAIF) | Graphonomous, PULSE, PRISM, Delegatic servers | live_deployed |
| Agent ↔ agent | A2A (AAIF, Aug 2026) | a2atraffic.com serves a signed Agent Card and answers A2A JSON | live_deployed |
| Skills | SKILL.md (agentskills.io) | ampersand-plugins; Workbench emits SkillBundles with proof gates | live_local |
| Composition | none industry-wide | [&] Protocol + CC2 compose laws (101 enforced) | in_tree |
| Governance | OWASP ASI · NIST AI RMF · EU AI Act | box-and-box (210+3 laws), OS-006 kernel | live_deployed |
| Loop declaration | none | PULSE manifests | live_deployed |
| Evidence | none (ISO 42001 A.6.1.6 wants it) | TRAAVIIS/trvs receipts · CLAIM_LEDGER | live_local |
Four of seven rows have no industry standard. Those are where a small portfolio can define terms; they are also where nobody is yet asking for one.
04 · Capital, and what it is not funding
Global venture funding reached about $510 billion in the first half of 2026, more than all of 2025, with AI taking over 70% of it (roughly 80% in Q1). OpenAI and Anthropic alone absorbed $217 billion, 43% of every startup dollar in the half. At seed, Carta's Q1 median was $3.2M raised on a $24M post-money, with AI companies carrying roughly a 42% premium over non-AI peers.
Underneath that, the shape is a barbell. Solo founders were 63% of new startups in 2026 by one payments-processor count, and “one-person company” is a normal thing to say about an infrastructure business. But fewer than 5% of the 12,000-odd public MCP servers had monetised by March 2026, and those that had mostly cleared $500 to $3,000 a month.
Two readings follow for this portfolio. First, its comparables have been funded at the paper stage: Letta at a research prototype, Mem0 before its graph tier. A working kernel with derived laws, an operating system image and a memory server on npm is more than several of those pitches carried. Second, that observation has been true since May and has changed nothing, because a term sheet is written against traction, not against a law count. Which returns to the review's finding: the missing input is one outsider, not more code.
| Datum | Value | Source |
|---|---|---|
| Global VC, H1 2026 | $510B | Crunchbase |
| AI share, Q1 2026 | ~80% · $242B | Crunchbase |
| OpenAI + Anthropic, H1 | $217B · 43% | Crunchbase |
| Seed median (all), Q1 | $3.2M on $24M post | Carta |
| AI seed premium | ~42% | Carta / Causo |
| Agentic-AI projects cancelled by 2027 | >40% | Gartner, Jun 2025 |
| “Real” agentic vendors | ~130 | Gartner |
| MCP servers monetised | <5% of ~12,000 | trade press |
| Typical MCP server revenue | $500–3,000 / mo | trade press |
| Cloudflare R2 storage | $0.015 / GB-mo · $0 egress | Cloudflare |
Do not plan revenue from MCP registry listings. Plan it from the storage SKU ([World] Cloud's cost basis is about $1.35 per 100 GB per month at R2 rates, with zero egress) and from certification, which is the one thing on this page a compliance officer can be told to buy.
05 · The memory market, close up
The competitor set is small, funded, open source, and converging. The table is what each one is betting on, and the last column is the thing this portfolio would have to show to matter.
| System | Bet | Funding | LongMemEval | Self-hostable | What Graphonomous has that it lacks |
|---|---|---|---|---|---|
| Mem0 | Extraction pipeline, multi-store fan-out; the community default (~51K stars) | $24M Series A | 94.4% (own report) | Yes; graph tier paid | A machine-checked cyclicity invariant; goals as graph objects |
| Letta | MemGPT hierarchy as an agent-as-a-server | $10M seed | LOCOMO-focused | Yes | Embedded SQLite; no server process to run |
| Zep / Graphiti | Temporal knowledge graph | $3.3M | 71.2% (GPT-4o judge) | Graphiti OSS, needs Neo4j | Consolidation with causal outcome attribution |
| Cognee | Ontology-aware graph from anything | $7.5M seed | no public F1 | Yes | κ-routed deliberation; PULSE manifest |
| Supermemory | Free local single-machine server since June 2026; engine closed | funded | — | Binary only | Open engine; the whole loop declared |
| Mastra OM · OMEGA | Observational compression; local-first layer | — | 94.87% · 95.4% | varies | Nothing on accuracy. This is the point. |
| Graphonomous | Continual-learning graph, κ invariant, goal coverage, edge-sized | $0 | 92.6% QA proxy, local reader | Yes, npm | — |
The status file records that topology on versus off on this benchmark is +0.3 points, and that no flat-RAG baseline was run. So the thing Graphonomous is for has not yet been shown to earn its keep where it is easiest to measure. The honest research position is that LongMemEval does not exercise cyclic knowledge, and the honest engineering position is that a benchmark which does needs to be published before the invariant is a selling point rather than a proof.
Where the category is going: the December 2025 survey “Memory in the Age of AI Agents” calls memory a first-class primitive; the 2026 papers that matter (NeSyC, trainable graph memory, ALMA) put a symbolic, checkable layer over the neural one; the May 2026 “replay divergence” preprint names the failure the compliance market is now paying for. All three point at provenance and replay, which is where the portfolio's instruments are, and away from the accuracy race, which is where its marketing was.
06 · A valuation ladder keyed to evidence rungs
The May page put the portfolio at $1.04M “realistic” on domains alone. That number had no method a reader could reject. This one does: the portfolio is worth what its highest occupied rung makes sellable, and each rung has comparables.
Bands are estimates by the method stated, not appraisals, and they widen to the right on purpose. The portfolio occupies the middle rung today. The step from the middle rung to the next one is the cheapest in the ladder and the only one that has not been attempted.
The 2026 aftermarket rewards single words (ai.com $70M, club.com $10M) and one-word .ai names (bot.ai $1.2M, neural.ai $1.75M). Compound coinages such as agentromatic or deliberatic are brands, not keywords; they are worth what stands on them, which is exactly why the ladder above prices the standing rather than the name. The two exceptions are computedriven.com and webhost.systems, which read as generic categories, and fleetprompt.com and specprompt.com, which carry a live keyword.
07 · The domains, one card each
Bare-value ranges are the aftermarket estimate for the name alone; “stands on it” is the rung the artifact behind the name has reached, from the review. Ordered by what is built, not by what the name is worth.
*.ampersand.json the CLI validates.Sum of bare ranges: roughly $60K to $180K, which is the bottom rung of the ladder above. Two names are not counted: toolboxhvac.com and brokenrecord.studio are outside the portfolio thesis.
08 · Strategy
AgenTroMatic, Deliberatic, GeoFleetic and TickTickClock are complete specs with no code and no near-term consumer. Parking them by name in the status file costs nothing, removes four rows from the scope grade in the review, and keeps the names, which are worth more than the specs. Un-park any of them the day a shipped product needs the primitive.
Everything on this page prices higher one rung to the right, and the step to that rung is the cheapest in the ladder: hand one outsider the docs atlas and one task, or get one stranger to reproduce one receipt. It has been runnable since 2026-08-11. Until it runs, the market research above describes a market the portfolio is beside, not in.
09 · Sources
Tags: analyst a named research firm · primary the organisation's own publication · survey a stated sample and method · vendor a competitor's own claim · press a named outlet · trade a trade blog, used where nothing better exists and labelled so.
Where a figure appears only in trade press (the 17% replay rate; the “55% of inference at the edge” family of claims), it is quoted with that label and should be weighted accordingly. No figure on this page was carried over from the May version without being re-found.