Every Company Is Building Agents That Can't Work Together - Justin Schroeder, Standard Agents
Build narrow, task-specific agents with explicit tools and a defined model, then compose them through a shared runtime standard.
- Keep agents focused on one business process and limit them to the tools that process requires.
- Evaluate each agent on its assigned task; pin a model that performs well rather than swapping models indiscriminately.
- Use a shared agent specification and runtime to package, reuse, and compose agents across systems.
The detailed walkthrough and live demo show how the proposed spec works in practice. · Best for: Teams building business-process agents that need controlled execution and interoperability.
Justin Schroeder argues that businesses need narrowly scoped agents for critical processes, rather than relying only on general-purpose models with tools and skills. He presents the Standard Agent Specification as a way to define, package, run, and compose those agents, then demonstrates generating and inspecting a warranty-claims agent.
Why businesses build agents0:00
- Businesses rely on important, organization-specific processes such as warranty claims, scheduling, financial aid, and billing disputes.
- For these processes, businesses need control, reliable repeatable execution, and the ability to preserve their own tone and practices.
Scope, trust, and efficacy3:32
- Limit an agent to a clear workflow and specific tools so it cannot take unrelated actions.
- Treat a complete task, such as handling a warranty claim, as an agent when it requires multiple turns.
- Evaluate agents on whether they complete their assigned task well; a smaller model may work well when paired with a narrow task.
Why a common agent standard8:45
- Agents involve more than a model and prompt: orchestration, parallel tools, error recovery, and hosting also matter.
- MCP, skills, and other standards address surrounding pieces, but the speaker says agents lacked a standard runtime.
- The proposed standard aims to make agents consistent, portable, and composable without rebuilding each harness.
Standard Agent Specification design13:29
- The specification defines agents as software harnesses around model outputs and emphasizes vendor independence, domain focus, model pairing, composition, and portability.
- Agents are written as TypeScript packages, with explicit models, prompts, tools, subagents, hooks, effects, and threads.
- Tools can be code, prompts, or other agents; threads can provide storage, code execution, decision models, and endpoints.
Runtime and builder26:52
- A runtime implements the specification's behavior while choosing its own underlying storage and infrastructure.
- The speaker describes Standard Agent Builder as a public runtime that can run locally or be deployed to the speaker's cloud.
- The specification is open, and the speaker invites others to build compatible runtimes and agents.
Warranty-agent demonstration29:53
- Claude Code generated a battery warranty agent from a PDF process description; the speaker reports it took 5 minutes and 21 seconds.
- The agent had a specific prompt, model, and explicit tools, and the builder displayed tool calls, model details, costs, tokens, and request data.
- The demo showed a web chatbot interface and described exposing the agent through HTTP, WebSocket, or MCP.
Step by step
- Choose one specific business process and document its rules, as with the warranty workflow PDF in the demo.
- Define a focused agent with a specific model, prompt, and only the tools needed for that process.
- Run evaluations against the task and inspect tool calls and model requests to verify performance.
- Package the agent for a compatible runtime, then compose it with other agents or expose it through an interface or MCP.
Tools & resources mentioned
- Standard Agent Specification (standardagentspec.org) — open specification for defining portable, composable agents.
- Standard Agent Builder — runtime and admin interface shown in the demo for running and inspecting agents.
- TypeScript — language used by the specification for agent packages and their components.
- MCP — discussed as a tool-distribution standard and as a possible way to expose an agent.
- Claude Code — used in the demo to generate the battery warranty agent from a process PDF.
Notes are AI-generated from the video's transcript and may miss nuance.