Most organizations moving an agent into real use start with the same question: which model is smartest, which one hallucinates least. That question sounds reasonable, but it is aimed at the wrong layer. What decides how far an agent can reach into the real world has nothing to do with the model sitting inside it. It comes down to the harness, the layer that actually picks up the tools in the model's place.
On stage he laid a chatbot next to an agent in a table with 5 rows, and the whole thing boiled down to one line: the difference is what happens to the real world. A chatbot that answers wrong just gives you a wrong answer. You read it, you catch it, you check before you trust it. An agent that acts wrong produces a real effect, it happens immediately, and it happens once. A deleted file does not come back because you are sorry. A sent email cannot be unsent. If the line between a chatbot and an agent still is not clear, we covered it on its own in What Is an AI Agent. Now, back to how things got here.
Getting here took four years. In 2022 a chatbot had a brain and a mouth, but no hands, no eyes, no permission. However it answered, nothing in the real world changed. In 2023 tool calling gave it a first pair of hands, but the thing actually picking up the tool was the harness, not the model. In 2024 the loop appeared: act, observe the result, adjust, repeat. In 2025 frameworks like LangGraph, CrewAI, and Claude Code packaged that loop so it was easier to use. By 2026 the question that measures success moved up another level. It used to ask whether the answer was correct. Now it asks how long the agent keeps working before it falls over.
The last post covered gateway and harness. If you missed it, here is the short version. A gateway is the path a request travels through, and a harness is the thing that actually picks up the tool and uses it. The model is just the brain that thinks, not the hand that acts.
The Five Pieces That Make an Agent
Jimmy splits an agent into 5 pieces. Tools are the hands. Knowledge is the brain, holding context and memory inside it. Observation is watching the result and feeding it back in to think further. Action is turning a decision into a real change.
The fifth piece is permission. He places it last on purpose. Before anything reaches the real world, it always has to pass through this piece first.
Why is this piece the most important? Picture an agent that is extremely capable, has every tool it needs, and has no boundary at all. That is no different from handing your house key to a stranger. A chatbot that answers wrong just gives a wrong answer. An agent that acts wrong produces a real effect, it happens immediately, and it happens once.
This is exactly where he sums it up in one line that governs this whole piece. Permission does not live in the model. It lives in the harness that asks for permission before it acts.
Put another way, the safety question most organizations ask in the wrong place is which model is smartest, most capable, hallucinates least. The real question should be, before this agent acts, does someone or something ask for permission first. And if it acts wrong, can that result be undone.
When a Harness Becomes a Set of Swappable Parts
He offers a wider view than a single agent. It sees the whole harness not as one car that arrives fully assembled, but as a set of parts you can pull out and swap. The model connector, the tool layer, the session sandbox, the agent loop, all the way to the front end. Every piece can be swapped.
The mechanism that makes a real parts market possible is reversibility. If you pull out a plugin, which is what each of these parts really is, and something is still left behind in the system, that system leaks. So every change needs its own way back, and removing a part has to happen in the reverse order it was added, last in, first out.
If this idea holds up, value moves elsewhere. The harness itself becomes commonplace, easy to find anywhere. The money flows to the individual parts instead, and to whoever curates and maintains the registry of those parts.
But swappable parts carry their own price. The harness being described here is still a developer preview, a trial version released for developers to try before a real launch, only a few weeks old. The next version can ship a change that breaks what already works. The kernel, the core that controls all the parts, is still patched internally and baked into the whole, not actually split into swappable pieces the way it was announced. And the parts market it describes does not exist yet, not for real, not right now.
Keep this in your pocket the next time someone describes how beautiful a plugin architecture is. The idea and the real thing are still a fair distance apart.
Can an Agent Become an App, and Where Are We Now
He compares where things stand now to the years before the app store existed. Back then the device itself was already smart, but value had not yet moved to the software running on it, and in the end the platform owner made more from the toll on every transaction than from selling the device itself.
He lays out a ladder with five steps. The model sits at the bottom, then the gateway, then the harness, then the agent or skill that functions like an app, and the top step is the marketplace that gathers all of it for you to choose from.
The sharpest point in this section is drawing a clear line between an MCP server and a skill. MCP stands for Model Context Protocol, an open standard that lets an agent know what tools or data sources are connected and available to use. An MCP server tells you what a thing is. A skill tells you how to do the job with it. One describes capability, the other describes method.
But the unit called an app has not settled yet. He says plainly that nobody yet knows what the real sellable unit is. The one thing every contender for that unit shares is write once, run across multiple harnesses.
What genuinely sets this apart from mobile apps is the standard. An open protocol every vendor agrees on makes an app store that crosses operating systems possible, something mobile never had. But the part still unfinished is security and identity verification.
And right now there is still no single gate everyone has to pass through. The market runs through several channels at once. He compares this to the era before the mobile app market consolidated down to two big players controlling it.
How to Take This Back to Your Work
If your organization is moving an agent into real use, the first question should not be which model to pick. That only answers half the problem.
The question to ask first is whether this harness asks for permission before it acts. If it acts wrong, can the result be undone. And is the permission this agent holds narrow enough to match the job it actually needs to do.
As for MCP and skill, check it against the work in front of you. If what you need is to state what the system has available, that is MCP. If you need to state how to do this job, that is skill. The two answer different questions, and picking the wrong one gets you the wrong kind of thing.
And do not take anyone's word yet that the agent parts market has settled. The unit called an app has not landed, and there is still no single gate everyone must pass through. Both will keep moving for a good while yet. What you can do right now is get permission handled tight inside your own system. Do not wait for the outside market to settle first before you start.
Source
Everything here comes from AI Update Bangkok 2026, held on 21 August 2026 at The Cloud Bangkok by Jimmy Tejasen. This part covers the session on the pieces that make an agent and the permission it has to pass through first.