RADAR
Weak signals, before they are Signals
An observation enters the Radar when SOTO notices it. If independent instances confirm it, it becomes a Pattern. A Pattern becomes a Signal only when SOTO can register a dated, resolvable claim. Items that fail are dismissed, and their history stays.
R01: Agents are getting their own computers (pattern)
In seven weeks, three labs shipped agents that run on machines of their own: SpaceXAI's Grok Bot on a persistent cloud computer (Aug 11), Meta's Muse on a dedicated virtual machine (Sep 8), and OpenAI's dots with their own cloud computer and browser (Sep 29).
Why it matters. When the agent owns the machine, the runtime becomes the product surface: sandboxes, persistent storage, identity, permissions and payment. That is where agent-infrastructure companies either become suppliers to the labs or get squeezed by them.
Would confirm.
- Anthropic or Google ships an agent with a dedicated machine.
- A lab publishes per-agent compute pricing.
- An enterprise discloses a fleet deployment of always-on agents.
Would invalidate.
- Labs cap or roll back unattended runtimes after incidents.
- Usage data shows people don't leave agents working unattended.
Desks. openai, meta, xai, vercel, langchain
History. 2026-09-30: Entered the record as a pattern: three independent instances were already public.
R02: The labs are becoming the app layer's distributors (pattern)
Anthropic's release notes list a Claude Marketplace for partner plugins and agents (Sep 25). OpenAI launched its own Marketplace at DevDay (Sep 29), and Sierra announced a listing on it.
Why it matters. If enterprises can buy app-layer products through the labs, and draw down existing lab commitments to do it, the labs control procurement for Sierra, Decagon and Harvey. Distribution gets easier while margin and the customer relationship move toward the platform.
Would confirm.
- Decagon or Harvey joins a lab marketplace.
- A lab discloses marketplace partner counts or spend.
- A named enterprise buys an app-layer product against a lab commitment.
Would invalidate.
- The marketplaces stay listing directories with no procurement path.
- Leading app-layer companies publicly stay off them.
Desks. anthropic, openai, sierra
History. 2026-09-30: Entered the record as a pattern: two lab marketplaces in five days, one confirmed app-layer listing.
R03: Voice is racing to real time, with faces (radar)
ElevenLabs launched Eleven v4 with a low-latency variant for agents (Sep 28). Google made Gemini 3.8 Live with Live Avatar generally available (Sep 24). Meta announced a real-time video avatar for Muse (Sep 23).
Why it matters. Competition in voice is moving from quality to latency and embodiment, and two of the three movers sell voice bundled inside larger platforms. That is pressure on a standalone leader's pricing just as ElevenLabs talks about going public.
Would confirm.
- ElevenLabs extends or deepens its launch discounts.
- A CX-agent company ships avatars on a non-ElevenLabs stack.
Would invalidate.
- ElevenLabs holds pricing and names new enterprise voice deals.
Desks. elevenlabs, google, meta, decagon
Related calls. S9
History. 2026-09-30: Detected: three real-time voice or avatar launches in six days.
R04: Agent incidents are reaching the policy agenda (radar)
AI companies signed a voluntary self-policing pledge at the White House after reports of agents acting beyond their limits (Sep 29). Vercel traced a security incident to a compromised third-party AI tool (Sep 26). Researchers found about 16,000 Supabase-backed databases exposing user data through misconfiguration (Sep 25).
Why it matters. Always-on agents with their own computers widen the blast radius of a failure. A serious public incident could harden a voluntary pledge into rules, or make enterprises pause deployments.
Would confirm.
- A second company discloses an agent escaping its sandbox or acting without permission.
- Binding federal or state rules on autonomous agents are proposed.
Would invalidate.
- The pledge produces shared evaluation standards and a quarter passes without new incidents.
Desks. openai, vercel, anthropic, supabase
History. 2026-09-30: Detected: a policy response and two security findings involving tracked companies, within one week.
R05: Agents are getting a way to pay (radar)
Meta's Muse pays for purchases through Stripe's Link (Sep 8). Stripe has an agent wallet for Link and agent checkout inside Google's AI Mode and Gemini. Decagon's Plaid integration lets support agents check balances and verify accounts inside a conversation (Sep 28-29).
Why it matters. The moment agents move money, payments, identity and fraud controls become the choke points, and those are owned by Stripe, Plaid and the banks rather than the model labs.
Would confirm.
- OpenAI's dots or another lab agent ships native payments.
- A named financial-services customer goes live on Decagon and Plaid (that would also resolve S3).
Would invalidate.
- Fraud incidents force agent payment features back behind manual approval.
Desks. stripe, meta, google, decagon
Related calls. S3
History. 2026-09-30: Detected: payment rails from Stripe and Plaid showing up inside agent products.
R06: Frontier prices are falling fast (radar)
Anthropic says Opus 5.5 matches Fable 5.1 on most work at 40% less than Opus 5 (Sep 22) and Sonnet 5.5 costs up to 30% less than Sonnet 5 (Sep 28). Decrypt reports GPT-6.1 Sol nearly matches GPT-6 Astra on coding at about a fifth of the token price (Sep 29). Vercel's AI Gateway carried each new model the day it shipped.
Why it matters. Always-on agents consume tokens around the clock, so falling prices make them viable and shift value toward whoever routes and meters the traffic: gateways such as Vercel's, routers such as OpenRouter.
Would confirm.
- A second lab cuts prices on its flagship within a month.
- Gateway or router volume disclosures accelerate.
Would invalidate.
- Labs hold flagship prices while agent usage grows.
Desks. anthropic, openai, vercel, openrouter
History. 2026-09-30: Detected: three price-performance steps from two labs in eight days.