Agents are deterministic software that harnesses the non-deterministic results produced by LLMs in pursuit of a business objective.General-purpose agents, like ChatGPT, have broad but limited utility. They cannot power the internal workings of a business without exposure to custom external tooling that interfaces with a business's surface area. The answer isn't Model Context Protocol (MCP) it's domain specific agents (DSAs). Software engineers already use DSAs: coding agents. Claude Code, Cursor, or Opencode (custom, domain-specific agents) are all examples of agents that are more useful for software engineers than ChatGPT or Claude — despite using the same models. The truth is purpose-built agents yield far better results in their given domain than general-purpose agents. Despite the religious fervor of those proclaiming AGI is close at hand domain-specific agents (DSAs) are here to stay. Standard Agents exists to answer the obvious next question: How do you create a domain-specific agent? ## How? The Standard Agents Specification defines the minimal set of parts required to create any agent:
The LLM provider and model configuration that powers the agent's intelligence.
System instructions that define the agent's personality, capabilities, and behavior.
Constraints and guidelines that govern how the agent operates and responds.
Functions the agent can call to interact with external systems and perform actions.
Lifecycle callbacks that run before or after agent actions for customization.
HTTP interfaces that expose the agent's capabilities to external applications.
A portable specification needs a baseline runtime that can execute almost anywhere. Without that baseline, implementations tend to depend on full server stacks (as in many MCP deployments). JavaScript is the most widely supported deployment target and can be sandboxed effectively, so it is a practical foundation. TypeScript then provides compile-time guarantees for both developers and LLM-authored code. Rather than maintaining custom validators, standard typechecking provides a simpler and more reliable contract.
MCP typically exposes one general agent to a broad tool surface, with limited tuning between model behavior, prompting strategy, and tool design. As the tool surface expands, quality and reliability often decline. Standard Agents takes a compositional approach: many focused agents, each tuned as a coherent unit across model, prompt, hooks, and tools, then orchestrated together. We did not land on the moon by giving one person every tool. We built teams of specialists.
Skills add useful context, and they can improve outcomes, but they still depend on a large general model to interpret and solve problems at runtime. This creates a similar scaling challenge to MCP: more context does not automatically produce better composition. Standard Agents composes complete, task-focused agent units rather than only adding instructions. We did not land on the moon by giving one person more documentation. We built coordinated teams.
It is an open source specification. Anyone can build or use a compatible runtime to execute any compliant agent. AgentBuilder is one such implementation.
A handoff reassigns control of the same thread to another ai_human agent. A subagent creates a new child thread using a dual_ai agent.
Resolution is layered: thread, user account, instance, agent defaults, then prompt defaults. Scoped variables can block parent inheritance for the branches that declare them.
Immediate tools run as soon as the prompt becomes active, before the first model step. This includes immediate subagent tool entries when they are enabled.
Yes. Optional subagent branches are controlled by environment flags (for example true, 1, or yes) and can be enabled or disabled per thread instance.