The system
How the system worksThe DirectoryDispatchThe LedgerAI modelsAI agentsYour installationThe appThe research libraryWhat changes for your businessUse it from your softwareEverything it can do
Messaging
iMessageWhatsApp
Ads and leads
AdsCreativeLeads
Devices
Devices
Industries
RealtorsRestaurantsHome services
Explore
How it worksPermissions and limitsCompareSet upQuestionsText me

The parts you can't buy yet

Take this build apart and price the components against the market. Three of its organs now map to funded companies. One maps to nothing you can buy — only to papers. That asymmetry is the most useful comparative fact on this site, so here

The parts you can't buy yet

Take this build apart and price the components against the market. Three of its organs now map to funded companies. One maps to nothing you can buy — only to papers. That asymmetry is the most useful comparative fact on this site, so here is the part-by-part accounting.

The text-message surface is now a $300 million company

The build answers natural language over iMessage: text the number, the kernel routes it, the reply carries real tool output. In April 2026 TechCrunch profiled Poke, from The Interaction Company of California — an agent that does daily planning, calendar, smart home, and photo edits "all via text message," on iMessage, SMS, Telegram, and WhatsApp. It raised $10 million on top of a $15 million seed at a $300 million post-money valuation, and in June became the first AI agent approved on Apple's Messages for Business.

So the surface bet was right, priced by someone else's investors. The difference in kind: Poke ships a consumer product with a fixed skill list; the build's texting surface fronts the same 887-row directory the API sees. One is a product, the other is an operating surface. Both prove texting is a first-class agent interface, not a demo.

Memory became a company too

Letta — the MemGPT lineage, 23,000+ GitHub stars by mid-2026 — treats "context windows as a constrained memory resource" and moves data between in-context memory and archival storage the way an OS pages between RAM and disk. Their framing is the right one: "designing an agent's memory is essentially context engineering."

The build's answer to the same problem is less clever and more literal: state files read at session start, an append-only cursor per project, and a ledger of every payload ever exchanged. No paging illusion. The comparison cuts both ways — Letta's agents self-edit memory inside the loop, which the build doesn't do; the build's memory is a flat auditable record, which a self-editing memory can't be. Pick your failure mode: forgetting versus never being able to prove what you knew.

Tracing became a dozen companies

Agent observability in 2026 is a settled product category. Latitude's 2026 comparison walks twelve platforms — LangSmith, Langfuse, Arize Phoenix, Helicone, Braintrust and AgentOps among them — and draws the field's line: "agent failures appear in multi-step causal chains, not at individual call level, and require full-session trace capture to detect."

The build's ledger records the same events with one structural difference: those platforms trace for debugging, and the trace is operational data you can edit or expire. The build's protocol layer hash-chains claim-changing events so anyone can recompute the chain and get valid-or-not. Trace answers "what happened?"; chain answers "can you prove what happened?" The market has productized the first question only.

The part that is still only papers

Authority is the gap. The build delegates by minting a capability URL scoped to one row or tier, with a TTL, a maximum use count, a stated purpose, and a risk ceiling — hand a model a link that can do exactly one thing for ten minutes, then dies.

The 2026 literature is circling exactly this. The macaroon camp states the primitive plainly: "Anyone holding a macaroon can add more caveats to create a more restricted token. This is called attenuation, and it's the foundation of safe delegation." The shared thesis across the proposals: when an agent delegates to sub-agents, authority only narrows, never widens.

What actually ships in mainstream stacks is OAuth scopes and long-lived API keys — identity-based, not capability-based. MCP's own auth spec is OAuth-shaped. No major agent platform today mints attenuating, expiring, purpose-carrying capabilities as its normal grant.

The scorecard

Reading the four comparisons together: surface — market caught up, and validated it. Memory — market went a different, fancier direction. Tracing — market matured past the build in tooling, short of it in tamper-evidence. Authority — market still in preprints. And the honest reverse column: the funded versions have eval suites, on-call teams, SLAs, and thousands of users hardening them; a single-operator kernel has one operator and a ledger. The gap that matters is which column closes first — products growing proof layers, or this build growing scale. Watch the capability-token layer; whoever ships it as a product turns the last paper column into a price.

References (5)
  1. Poke makes using AI agents as easy as sending a text (TechCrunch)“all via text message”
  2. Agent Memory: How to Build Agents That Learn and Remember (Letta)“Designing an agent's memory is essentially context engineering: determining which tokens enter the context window and how they're organized.”
  3. Best AI Agent Observability Tools in 2026 (Latitude)“Agent failures appear in multi-step causal chains, not at individual call level, and require full-session trace capture to detect.”
  4. Macaroon Tokens vs API Keys for AI Agents (dev.to)“Anyone holding a macaroon can add more caveats to create a more restricted token. This is called attenuation, and it's the foundation of safe delegation.”
  5. Agent Authentication & Delegated Access (Zylos Research, Apr 2026)“93% of AI agent projects are still using unscoped API keys, and 74% of respondents said their agents end up with more access than they actually need.”