The core thesis: Most AI agents today are LLM wrappers. They reason in text but depend on brittle scripts or APIs to act. Large Action Models (LAMs) take a different approach — they model action sequences, not token sequences.
LLMs vs LAMs
| LLM | LAM |
|---|---|
| Models token sequences | Models action sequences |
| Output: words | Output: clicks, keystrokes, navigation |
| Requires APIs/schemas | Operates on raw UIs |
| Brittle to interface changes | Adapts to UI drift |
The Action Loop
LAMs execute through a continuous loop:
- Screen perception — visual + DOM signals
- Policy inference — next best action
- Actuation — mouse/keyboard events
- State verification — confirm action succeeded
- Iterate — recovery from partial failure
This enables handling UI drift — when a website updates, the agent can recover and find alternative paths rather than breaking entirely.
The Action Tree
As agents operate, their behavior is logged into an Action Tree:
- UI states as nodes
- Valid transitions as edges
- Successful paths marked
- Failure cases logged
Over time, this becomes a reusable action prior across applications — the agent "learns" common UI patterns.
Actionist Runtime
"Runtime for AI operators":
- Local desktop execution
- Cloud VPCs for scale
- Replayable agent history
- Full observability
Agents don't run in prompts. They run in environments.
Training Data
Instead of scraping or synthetic demos, training comes from real human interactions captured via controlled tooling. This grounds the model in how software is actually used, not how developers think it's used.
ActionFi: Tokenized Incentives
Action Model introduces a token layer where:
- Tasks are verifiable
- Execution is auditable
- Rewards tie to completed actions, not promises
The token represents fractional stake in the AI — economic rights, governance rights, participation in growth.
The Broader Thesis
APIs won't cover the long tail of software. Humans already did that work through interfaces. LAMs learn from that surface area and turn it into executable intelligence.
If LLMs are about reasoning in language, LAMs are about operating in the world.
Status
- Pre-launch building phase
- Browser extension available for data contribution
- Creator Circle recruiting
- ActionFi launching "in a few weeks"
Source: @dami_Rtetyl3 thread + @ActionModelAI tweets, synthesized January 30, 2026.