The Next Evolution of Agents Will Come from Smarter Tools
TL;DR
- Like SOA before it, today’s smart pipeline, dumb edge structure may approach its limits and shift toward dumb pipeline, smart edge, where the tools themselves become intelligent.
Introduction
What people building or using agents do most often these days is ask what else they can attach to the main agent. They add more Skills, install plugins, and connect MCP servers. The main context keeps getting heavier, while tools are carved into forms that are convenient for that context to call.
Viewed from one step back, this picture starts to look familiar: smart pipeline, dumb edge. The central pipeline owns all the intelligence, and the tools at the edges are assumed to be better when they are simpler. It resembles a shape we encountered once before during the SOA era.
1. SOA déjà vu: today is smart pipeline, dumb edge
Anyone who remembers the Enterprise Service Bus, or ESB, from the Service-Oriented Architecture, or SOA, era may feel the same resemblance. A central bus owns all routing, transformation, orchestration, and protocol mediation, while the services at the edges remain as thin and simple as possible. It is clean in theory. But real-world operations often return to the same problem: the moment the center absorbs every responsibility, that center easily becomes both the complexity and the bottleneck of the entire system.
The industry gradually moved in another direction. Messaging infrastructure such as Kafka is a good example. The pipeline becomes intentionally simple, or a dumb pipeline, while domain knowledge and decisions move to producers and consumers at both edges, creating smart edges. We have already experienced a transition in which the whole system can behave intelligently without an intelligent center.
Looking at today’s agent ecosystem, it sometimes feels as if we have returned to the stage before learning that lesson. Every tool hangs from the ESB called the main agent.
2. The next stage is dumb pipeline, smart edge
When agentic engineering matures further and moves into full-scale automation, might this structure naturally reverse?
If one main agent must keep track of dozens of dumb tools as it does today, the main agent’s context and reasoning capacity can easily become the upper bound of the whole system. As more tools are added, the main context grows heavier, reasoning slows, and throughput seems likely to approach its limit. I can imagine a curve similar to the way an ESB itself became the system bottleneck as it grew.
The picture could be quite different if each tool were a carefully designed vertical agent. The main agent would need to focus only on orchestration. Saying it can remain relatively dumb does not mean that the main agent becomes stupid. The nuance is closer to moving from CPU-bound to IO-bound. While specialized agents at the edges distribute and process the heavy reasoning, the main agent focuses on whom to call, when to call them, and how to combine their results. Throughput may have room for a step change.
For this picture to work, tools at the edges probably cannot remain “bundles of functions carved for the main agent’s convenience.” An edge needs to be closer to an agent that can make decisions within its own domain. It should manage its own context, perform its own validation, and avoid depending on the main agent’s intelligence. This is why I expect tools built around vertical agents to become increasingly common.
3. The same direction in physical AI
I think this perspective can also be applied to physical AI.
One major reason today’s robotics industry is converging on humanoids is, in my view, the ease of acquiring data. Human motion data is the most abundant, and the humanoid form fits environments built for people better than any other. Rather than being the most effective shape everywhere, a humanoid seems closer to the easiest shape to train at this point in time.
But as populations decline and urbanization continues, perhaps the asset whose value rises fastest will be space. My wife and I currently live together in a one-and-a-half-room apartment. Neither of us wants much, so it is not particularly uncomfortable, but small is still small. If we added even a highly capable humanoid, never mind the paths it would need to move through, I think I would feel suffocated just seeing it stand still. A humanoid is inherently a form factor that takes up considerable floor space and room to move.
Even so, the reason humanoids appear attractive seems fairly clear. Every tool in the home is extremely dumb. A refrigerator only stores food, a washing machine only washes clothes, and a gas stove only lights a flame. If we leave these dumb edges in place and try to make the home intelligent, we will probably need something like a general-purpose actuator that can operate every tool in place of a person. The most familiar form of that actuator is a humanoid.
This picture is essentially smart pipeline, dumb edge: a humanoid at the center plus dumb appliances at the edges.
4. A smart-edge home looks different
Once every appliance in the home can be controlled through an API, the picture may look quite different.
An adult-sized humanoid may no longer be necessary. If appliances can operate autonomously within their own domains, the central robot has less reason to imitate a person and do everything on their behalf. A robot about the size of a child, perhaps with arms that extend when needed, might be enough. It could handle a useful range of tasks while taking up very little space or obstructing anyone’s path.
Let us go one step further and imagine the appliances themselves becoming true smart edges.
- A kitchen appliance with robot arms and built-in recipes and cooking methods
- A washer-dryer with robot arms that identifies fabric automatically and handles washing, drying, and folding
- A refrigerator that tracks ingredients directly and automatically orders whatever is running low
If this happens, the orchestrator inside the home may be reduced to connecting the outputs of each smart edge to the person. It might carry laundry from the washer-dryer to the closet or food from the kitchen to the table. If that is only a little more convenient than walking around and doing it ourselves, perhaps that is enough.
There may be one difference from workplace automation: a home is a place where people live. In work, the direction of agentic engineering seems to be moving toward removing humans from the loop,1 but there is no particular need to do that at home. A person is present in that space and is the one who enjoys the result.
If the remaining work is merely walking around to collect the outputs of smart edges, that seems worth trading off against a humanoid that blocks the paths through the home while doing everything itself. Instead of giving up scarce space for a human-sized robot, fill the home with a small orchestrator and intelligent appliances. It is another way of buying the same automation with a different space cost.
5. The place for SaaS companies: headless SaaS
Bringing this trend back to software, I think it may also affect how SaaS companies position themselves.
Until now, SaaS has been closer to a model that owns the UI, workflow, integrations, and automation within its domain all at once. Its value came from letting users begin and finish their work inside the SaaS product’s screens. But if the center of orchestration moves away from a person’s screen and toward the main agent, might that value structure begin to shake as well?
I can imagine the main agent, or the small number of platform companies that build it, owning orchestration itself. Within that structure, the natural direction for each SaaS company seeking a meaningful place may be to become the smart edge specialized in its own domain. In other words, headless SaaS: a shift away from being a company that holds a UI and consumes user time, and toward being a company that supplies the main agent with an agentic tool that performs sophisticated work within its domain.
Instead of focusing on orchestration itself, focus on building a domain-specific tool that does genuinely intelligent work when the orchestrator calls it. The agent with the deepest context and most sophisticated harness in one area may become the default for that domain. It resembles the smart appliances described above.
6. From a developer’s perspective: domain knowledge and the Forward Deployed Engineer
Viewed again from a developer’s perspective, the work ahead may ultimately resemble building the smart edge for your own domain.
Everyone is likely to use the same models, and orchestration layers may also become increasingly standardized. In the end, the difference will probably come down to how deeply you understand the domain and how precisely you can transfer that understanding into code and tools. The depth of a smart edge ultimately comes from domain knowledge.
From this perspective, it also seems natural that AI companies have recently been hiring heavily for roles such as AI Deployment Engineer and Forward Deployed Engineer. Improving the model itself remains important, but this can also be read as a sign that the value of people who carry the model deep into a customer’s domain and shape it into that domain’s smart edge is rising with it. At the point where one general model meets a domain, the engineer who understands that domain best may soon hold the greatest leverage.
Conclusion
Today feels like the era of smart pipeline, dumb edge. The main agent carries all the intelligence, while tools are carved around the convenience of being called by that agent. As with SOA, once this structure passes a certain scale, the center itself may become the bottleneck.
That is why the next stage seems likely to move toward dumb pipeline, smart edge. The main agent stays light and IO-bound while focusing on orchestration, and tools at the edges work with their own intelligence inside their respective domains. In software, this could take the form of vertical agents and headless SaaS. In physical AI, a similar direction could appear as small orchestrators and smart appliances.
Time will ultimately answer which picture is more natural: one humanoid compensating for a dumb home, or everything in the home becoming intelligent until only a small orchestrator remains. Still, having already watched the industry move from SOA to Kafka once, perhaps I can vaguely guess which way the answer will lean.