Is there a standard for Large Action Models?

OpenLAM publishes an open reference for the Large Action Model category: the LAM core loop (plan → act → observe → learn), the Open Agent Protocol (OAP) for declaring and validating actions/intents, and open benchmarks. It is open source so the ecosystem can build on a shared definition.

What is the LAM core loop?

The LAM core loop is the execution cycle every Large Action Model agent runs: perceive the goal and state, plan the steps, act by calling a tool or API, observe the real result, and learn from it. Repeating this loop is what lets a LAM complete multi-step work autonomously.

What is the Open Agent Protocol (OAP)?

OAP is OpenLAM’s open way to declare the actions (intents) an agent can take, validate a planned action before it runs, and report results honestly. It makes capabilities explicit so an agent can ask for a missing tool instead of failing or hallucinating success.

The LAM Standard, Protocol & Whitepaper

Large Action Models are a category, and OpenLAM defines it in the open. This is the reference for what a LAM is: the core loop every agent runs, the Open Agent Protocol for declaring actions, the benchmarks we measure against, and the open whitepaper.

The LAM core loop

Every Large Action Model agent runs the same fundamental cycle. It is the defining mechanism of the category:

Perceive — load the goal, available tools, memory and current state.
Plan — decompose the goal into an ordered set of concrete, executable actions.
Act — execute one action: call a tool, an API, the browser, or delegate to another agent.
Observe — capture the real result and feed it back into the plan.
Learn — record what worked, so the workforce improves over time.

The Open Agent Protocol (OAP)

OAP is the contract layer that makes actions explicit and honest. An agent declares the intents(actions) it can perform; a planned action is validated before it runs; and results are reported truthfully. Crucially, when a goal needs a capability the workforce does not have, OAP lets the agent ask for the missing tool rather than fabricate success.

Why a protocol matters

Capabilities are discoverable and verifiable, not implicit.
Plans can be checked for feasibility before execution (no silent failures).
Tools and plugins — including any MCP server — plug into a common action surface.

Benchmarks

A standard needs measurement. OpenLAM evaluates the workforce on real task completion — did the goal get done, was the deliverable correct, did it need human correction — and uses those signals in its learning loop (and optional privacy-first federated learning) to promote only models that beat the current one on a held-out benchmark.

The open whitepaper

OpenLAM is open source (AGPL-3.0). The implementation of the LAM core loop, the OAP intent surface, and the evaluation harness are all in the public repository — the living whitepaper for the category.

Read the source on GitHub →

Learn the fundamentals

Start with What is a LAM?, see how it differs from LLMs and agents in LAM vs LLM vs Agent, or build one in How to build agents with OpenLAM.