The Standard Agent Specification is an open-source standard for creating portable, composable, domain-specific agents.

Start with the basics.

As a baseline every agent needs:

• A model to power it
• A system prompt to steer it
• Tools for it to describe actions
• And an ongoing conversation for context

Expand the capabilities.

Today agents rely on MCP servers and skills to do real-world work.

But there’s a problem…

Context isn’t free.

The larger the scope of your agent, the more expensive the model required to power it. Even the largest, most expensive frontier models eventually hit their limit.

Stacking every responsibility into a single entity is inheritance, and it inevitably breaks down.

But we already know the solution.

Effective engineering has always been about composition over inheritance.

So let’s scope our agents.

By focusing each agent on a specific domain, we will create many agents, each with its own specialty.

We can then leverage cheaper models, taking advantage of their jaggedness, and limit them to only the capabilities necessary for their work.

And compose them into effective teams.

If our agents are built to the same open standard, then coordination between them is easy.

What does a standard agent look like?

• Rules for limits like how many turns an agent gets.
• Hooks to inject, compact, and alter context.
• Sub-prompts to swap prompt and model for one action.
• Sub-agents to hand off tasks to their proper domain.
• The exact provider and model to guarantee efficacy.
• Storage for long-lived assets and memory.
• Code execution to (safely) run arbitrary code.

Agents are made of agents.

Standard agents are invoked by other agents like any other tool call. Each standard agent maintains its own complete isolated context.

Coding agents can produce standard agents quickly because every architecture and runtime decision is predecided in the open specification.

Sharing becomes easy.

Because the entire agent stack is defined in code, packaging and sharing agents is trivial.

Every standard agent is guaranteed to run in any spec-compliant runtime.

An open marketplace of agents can flourish.

A wide array of community-built agents suddenly becomes just an install command away.

Each agent executes exactly as its original author intended.

Agents can be installed, unpacked, modified, and republished at will, allowing you to adjust them to your specific needs.