Justin Schroeder of Standard Agents introduces a proposed open agent specification and a builder runtime at AI Engineer. His 35-minute talk ends with a recorded battery-warranty workflow.

Why package a narrow agent?

  • An HVAC warranty claim needs equipment lookups, photos, eligibility checks and a controlled submission path. A broadly equipped assistant has more ways to act outside that workflow.
  • Schroeder argues that businesses need integration, bounded authority and efficacy: whether the specific task gets done reliably.
  • A small agent can own several conversational turns and tools while being evaluated against one business process. Its results still depend on the chosen model and tests.

What the proposed standard contains

  • MCP standardizes tool connections; this proposal specifies the agent harness: model, prompts, tools, sub-agents, rules, hooks, scheduled effects and conversation state.
  • Agents are TypeScript packages; another compliant runtime can interpret them. The runtime itself can be written in another language if it executes the specified package.
  • The agent pins a model and can define a fallback. Code tools, model-backed prompt tools and whole sub-agents can be composed.
  • Hooks run at lifecycle points; scheduled effects handle later actions. A thread carries its own messages, key-value state, files and optional code-execution permissions.
  • Schroeder says a runtime may back the thread’s apparent filesystem with a database. Portability depends on implementations matching the contract, not on copying the same storage engine.

The warranty demonstration

  • In a sped-up recording, Claude Code reads a process PDF and creates a battery-warranty agent with narrowly named tools in 5 minutes 21 seconds.
  • The builder shows thread traces, model requests, tool choices, token use and a 94% cache-hit figure for the displayed request. The example ends with a replacement-shipping tool call.
  • A website chat connects to the agent through HTTP/WebSocket; Schroeder proposes exposing the same agent to another assistant through MCP.
  • Caveat: one recorded build and demo do not establish portability across independent runtimes or the warranty agent’s measured accuracy. The speaker explicitly calls the specification unfinished.

“A good agent is a good harness.” — Justin Schroeder, 13:53