Watch the "one all-purpose agent plus MCP" path get taken to its breaking point, then get a compositional alternative
The Future Is Domain-Specific Agents - Justin Schroeder, StandardAgents · Justin Schroeder
31 min total·Actually worth watching closely: ~2 min·2 must-watch clips
- 0:03 – 7:03Listen
Everyone is building agents, then getting stuck on integration
Opens with the Industrial Revolution analogy to explain why everyone is building their own agents this cycle, then lays out the reality: a robust agent takes careful loop orchestration, durable execution and a ton more, so most home-grown agents stop at demo quality and the industry has no agreed standard.
The hard part of building your own agent was never the model — it's the orchestration and reliability layer nobody wants to do.
Mostly industry observation and analogy delivered standing up; the screen adds nothing, so listening while you do something else is fine.▶ Jump to 0:03Speaker · Justin Schroeder - 7:03 – 9:20Skim
How much of MCP is actually supported
Rolling your own is too hard, so everyone turns to MCP — but he puts the client capability support up for comparison: tool calling is the only thing universally implemented, and MCP's other promised capabilities are largely unshipped.
In practice MCP has degraded into a tool distribution channel, and handing an agent tools alone isn't enough to get things done.
What matters is that table of checks and blanks on screen; scan how the columns fill in and the conclusion is obvious — you can skim the narration.▶ Jump to 7:03Speaker · Justin Schroeder - 9:20 – 14:06Listen
The ceiling of extension by inheritance
Frames the habit of piling more MCP servers and skills onto the same agent as inheritance in object-oriented terms: context keeps thickening, returns diminish, and once you load up on skills the agent gets substantially worse.
A monolithic agent's capability doesn't accumulate linearly — past a point it starts going backwards, and then it breaks down.
Pure argument here, carried by the inheritance-versus-composition thread; nothing on screen you need to see.▶ Jump to 9:20Speaker · Justin Schroeder - 14:06 – 16:50Skim
Swap in a team of specialists
The alternative: not one all-purpose agent but many domain-specific ones, each with its own message history, tools and agentic loop, with a coordinator above directing them in plain English — the way the specialists in mission control, each owning one station, worked together.
Composition over inheritance — every one is a complete, isolated agent, and they collaborate just by talking to each other.
The mission control frame plus his brain/dashboard/mouth breakdown is best watched; outside that image you can move quickly.▶ Jump to 14:06Speaker · Justin Schroeder - 16:50 – 21:09Listen
The tokens saved and the small models you can finally afford
The payoff, with numbers from practice: a sub-agent needs only a system message, a few tools and one incoming instruction rather than the entire conversation, which they regularly measure at over 80% token efficiency. Push narrow tasks to a cheap model and per-task cost differs by two orders of magnitude. Permissions also tighten to "only what's already approved," which enterprise IT is far happier with.
Shrink the context window to the minimum and small models become genuinely usable; the savings and the control come from the same architectural choice.
All the numbers are spoken outright and there's no chart to read cell by cell, so listening is enough.▶ Jump to 16:50Speaker · Justin Schroeder - 21:09 – 28:12Listen
The price inflection and a parts list for the ideal agent
First a counterintuitive claim: the consensus that intelligence keeps getting cheaper flipped this year — capability-adjusted token prices are up nearly 30% in six months, over 70% unadjusted. Then a part-by-part breakdown of what an ideal agent should have: three kinds of tools, hooks that inject information or fire side effects, rules that bound turns and validate calls, and a sandboxed filesystem of its own.
Unless a customer's lifetime value is enormous you can't put a big model in front of them — cutting cost went from optional to prerequisite.
The component list is laid out item by item in speech; following his order sticks better than watching the screen, which makes it good commute listening.▶ Jump to 21:09Speaker · Justin Schroeder - 28:12 – 30:34Listen
What multi-layer orchestration looks like, plus two predictions
What it looks like in practice: a top-level coordinator directs domain agents that each know their own APIs and hold their own credentials to get work done across systems, sub-agents can call sub-agents of their own, and context-heavy domains like compliance keep branching into dedicated specialists. Closes with a dated forecast.
The further down you split, the smaller each agent's context window — which is exactly why the whole architecture scales.
The ending is a spoken description of the hierarchy plus the predictions; dense, but it doesn't depend on the visuals, so listening covers it.▶ Jump to 28:12Speaker · Justin Schroeder