Sunil Venkataram
Back to garden
•
🌰 seedling
Claws - Persistent Looping Agents as App Replacement
type/noteaiagentsproduct

Claws — Persistent Looping Agents as App Replacement

The Dobby example

Andrej Karpathy describes a personal experiment he calls the “Dobby the elf claw”:

  • Started with a persistent agent loop on a laptop
  • Scanned the local network for smart home devices (Sonos speakers, lights, HVAC, pool controls, security cameras)
  • Found endpoints, reverse-engineered the APIs where documentation was thin
  • Built a command dashboard abstracting all of them
  • Natural language control: “Dobby, sleepy time” → lights dim, thermostat drops, security arms
  • Camera + vision model → detect FedEx truck at the door → WhatsApp message with the image

The result: six different apps collapsed into a single natural-language interface. The apps still exist underneath, but the user never touches them. The agent is the interface. The apps are backend.

Why this is a platform argument, not a parlor trick

The obvious read is that Karpathy built a fun personal gadget. The real claim is stronger:

Most consumer apps should expose APIs; agents should be the intelligent glue. The customer is increasingly an agent, not a human.

That reframing has teeth. If agents become the dominant consumers of consumer software APIs, then:

  • UI investment becomes diminishing-returns. Why build a beautiful mobile dashboard if the dashboard’s users will be other agents?
  • API quality becomes the battleground. Apps with clean, well-documented, semantically consistent APIs get composed into agent workflows. Apps with messy APIs get wrapped or ignored.
  • Pricing models need to change. Per-seat human pricing doesn’t make sense for a user who is a background agent running 24/7 on behalf of one human.
  • Integration lock-in loses power. If the agent is the glue, it doesn’t matter which calendar or which notes app you use — the agent bridges them.
  • Discovery gets weird. Humans don’t install apps they never see. Agent-to-agent discovery protocols (MCP, tool manifests, capability registries) become the new App Store.

Claws vs. traditional agents

A claw is specifically a persistent looping agent — it runs continuously in the background, maintains state across cycles, and responds to ambient signals (network events, camera feeds, calendar changes, sensor data). That’s different from:

  • One-shot agents: given a task, complete it, return. No persistence.
  • Chat agents: respond to human messages, go idle between turns.
  • Scheduled jobs: wake up on a timer, run a script, sleep. No intelligence between runs.

The claw sits in the gap: it’s always on, it reasons about its environment, and it can take autonomous action when conditions match. The closest predecessors are classic cybernetic control loops, but with an LLM at the reasoning step.

What breaks if this is right

  • The “app economy” as a zero-sum fight for home screen icons matters less than the API layer underneath.
  • Device-first and phone-first design competes with always-on-in-the-background-first design. Background-first may win for utility; device-first holds for media and social.
  • Privacy and trust boundaries get harder, not easier. A claw with access to your smart home, calendar, bank, and camera can also do those things without asking.

Where this is already happening

  • Smart home integration platforms (Home Assistant, SmartThings) are early-stage claws that lack the LLM reasoning step
  • Voice assistants (Alexa, Siri, Google Home) were supposed to be claws but stalled at single-command interfaces
  • MCP server ecosystems are the protocol substrate for claws to compose any API they can discover
  • Coding agents with filesystem + shell + browser access are the most mature claws in production today
Backlinks (6)
Connected Notes