Most AI assistants are designed to do almost anything. Justin Schroeder, co-founder of Standard Agents, asks whether that is the wrong shape for a business that needs one job done correctly every Tuesday. In this Agentic AI Foundation talk, he argues for many small, domain-specific agents: each owns a bounded process, uses only the tools it needs, and can be tested on the outcome the business actually cares about.
Do not ask one all-access assistant to run an entire company. Give a focused agent one recurring process, a limited tool set, a clear stopping point and a private acceptance test. Let an orchestrator coordinate specialists when a job crosses domains. This is Schroeder's architectural argument, not proof that every multi-agent setup is cheaper or safer by default.
Watch Justin Schroeder's Talk
Schroeder's AI Engineer speaker page gives more background and links to a separate, earlier talk on the same theme.
The Factory Analogy: One Engine, Too Many Belts
At 02:18, Schroeder compares today's broad assistant to an old factory's central steam engine. Every workstation drew power through belts and line shafts. The arrangement worked, but each job depended on one sprawling mechanism. His proposed replacement is a small power source at each workstation: an agent dedicated to a particular domain, with coordination only where needed.
His working definition of an agent is deterministic software using non-deterministic model results toward an objective. The model is one component of a harness that may also include prompts, tools, hooks, execution limits, storage and recovery behavior. Calling an LLM in a loop is easy; keeping a real process running correctly after errors, interruptions and changes is the engineering work.
The analogy is useful, but it should not be mistaken for a ban on skills or MCP. Schroeder explicitly says those can remain part of the system. The architectural question is where responsibility and authority live: in one general assistant with access to everything, or in smaller components whose scope can be inspected and tested.
Why a Business Might Prefer Specialists
| Reason | What Schroeder means | Test in a real operation |
|---|---|---|
| Integration | The agent fits an existing process, including its data, handoffs and stopping point. | Can it move a fleet-maintenance case from reported issue to approved repair and updated records without replacing the rest of the business? |
| Trust | Tools and permissions are bounded by the job, not merely by a prompt asking the model to behave. | Can a mail agent read and draft in Gmail while being technically unable to change production infrastructure? |
| Efficacy | The outcome can be evaluated repeatedly against the same process-specific expectations. | Does it classify urgency, preserve the evidence trail, request approval and close cases correctly across representative examples? |
The example at 07:05 is a heating and air-conditioning company. It already has lead capture, scheduling, customer service and many other working processes. A fleet-maintenance agent could instead tackle just one: receive a vehicle issue, assess priority, schedule a repair, update the record and return the vehicle to service. This is a proposed design example, not a documented deployment at that company.
What a Small-Agent System Looks Like
At 10:24, Schroeder contrasts a Gmail agent with a general assistant given a Gmail skill. His Gmail specialist knows the mail API and cannot wander into another domain because it lacks those tools. At 11:22, he places Gmail, Figma and Sheets specialists beneath an orchestrator that delegates work and assembles the results.
| Layer | Responsibility | Boundary to enforce |
|---|---|---|
| Orchestrator | Understand the request, select the right specialist and combine outputs. | Delegation does not automatically grant unrestricted credentials. |
| Domain agent | Own one process or integration, with its own instructions, model and context. | Give it only the tools, data and action limits that the process needs. |
| Deterministic controls | Validate inputs, log actions, retry failures and pause for approval. | Do not rely on a prompt alone to enforce a permission or spending limit. |
| Process evaluation | Replay representative cases and score accepted outcomes. | Measure errors and reviewer work, not just a model benchmark rank. |
A skill is not inherently unreliable, and a dedicated agent is not inherently secure. The dependable part comes from real access control, clear interfaces, safe failure behavior and tests. Multiple agents also introduce coordination cost and new failure modes, so a single simple automation may be enough for a small, deterministic job.
Efficacy, Jagged Models and Token Cost
Schroeder's strongest operational point arrives at 15:00: a bounded workflow is easier to evaluate than a universal assistant. For the HVAC example, a useful test set would include duplicate reports, missing vehicle IDs, an urgent safety issue, a vendor cancellation and an unapproved repair quote. Pass criteria can be explicit: correct priority, no unauthorized purchase, complete audit trail and timely handoff.
Models also have uneven strengths. In the talk, Schroeder shows one small model producing a 3D scene he prefers to a larger model's result. That is an illustration of jagged capability, not a general ranking. A narrow agent can be paired with a model that performs well enough on its own cases. Before switching models, replay the same cases and compare accepted outcomes, latency, token use and human correction time.
He reports large savings in Standard Agents' private tests and cites a striking model price comparison. Neither the private evaluation set nor its complete cost accounting is published in the talk. Treat those figures as claims to verify for your own workload. Token price alone does not capture retries, tool calls, orchestration overhead or review effort.
What the Standard Agent Specification Now Provides
The talk's 21:24 section describes an open-source specification and runtime as something Standard Agents hopes to publish. That status has changed: the official Standard Agent Specification site now lists v0.2.0. It sets out principles for vendor-agnostic, domain-specific, model-coupled, composable and portable agents. Its definitions cover providers, models, tools, prompts, agents, subagents, hooks, packaging and runtime responsibilities.
Those terms need care. Vendor-agnostic means the format is intended not to belong to one model company. Model-coupled means a particular agent can still specify the provider and model it was evaluated with; casually swapping that model may change behavior. Portable is the goal that a conforming package executes faithfully on conforming runtimes, not evidence that every runtime or every model interchange is equivalent.
The episode's description links to v0.1.0 documentation. For new work, begin with the current v0.2.0 overview instead. The runtime directory presently lists Standard Agents Builder as its available runtime and explicitly says listing is not a conformance certification. The specification is a published proposal and implementation path; broad independent interoperability is still to be demonstrated.
How to Test the Idea Without Rebuilding the Company
- Choose one recurring process. Draw its start, end, data sources, handoffs and exception paths. Fleet maintenance, invoice reconciliation or a support triage queue are better-defined starting points than “run operations.”
- Define the authority. Separate read, draft and write actions. Put spending, external messages and irreversible changes behind explicit approval or deterministic policy.
- Build a small evaluation set. Include ordinary cases and failures. Score completion, prohibited actions, evidence quality and how often a person has to intervene.
- Compare architectures. Run the same cases through a simple automation, a broad assistant with a skill and a dedicated agent. Include integration and maintenance work in the comparison.
- Version the entire setup. Record model, provider, prompts, tools, permissions, tests and release date. Upgrade only after the process-level checks pass.
Schroeder's Q&A closes on governance: an IT team needs visibility across its agent fleet and private evaluations before changing a model. A stable model need not be replaced just because a newer one exists, though deprecations, security issues or changed requirements can force a review.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | Introduction |
| 02:18 | The line-shaft era of agents |
| 03:08 | Domain-specific agents |
| 03:33 | What an agent is |
| 05:29 | Why businesses build agents |
| 07:05 | The HVAC fleet-maintenance example |
| 10:24 | A Gmail agent versus a Gmail skill |
| 11:22 | Specialists under an orchestrator |
| 12:04 | Engineering principles for agents |
| 15:00 | Evaluating one business process |
| 15:42 | Jagged model capabilities |
| 17:14 | Token spend and cheaper models |
| 18:53 | Why factories needed standards |
| 20:23 | What portable agents need |
| 21:24 | The proposed spec and runtime |
| 22:01 | Q&A: agent logic versus model behavior |
| 23:16 | Q&A: maintenance and governance |
Sources and Limits
This article uses the Agentic AI Foundation recording and the supplied transcript for Schroeder's argument; the speaker biography for context; and the official specification and runtime directory for the post-recording status update. We have not benchmarked a Standard Agent deployment or verified the talk's private cost claims. Architectural recommendations and historical analogies remain Schroeder's interpretation. Evidence cutoff: 16 November 2026.