Predictions about AI age badly, so this chapter is deliberately about directions rather than dates.
Predictions about AI age badly, so this chapter is deliberately about directions rather than dates. What follows is where the effort is currently going, and why each of these matters to somebody who now understands how these systems are put together.
Today's agents answer questions and book meetings. The direction of travel is towards systems that behave less like tools and more like colleagues.
Picture an agent that books your travel, rebooks when the flight is delayed, notifies the hotel and messages your client that you will be late. That is not one agent. That is a small coordinated team — and you now know how it would be built, roughly what it would cost, and which of those four actions should have stopped to ask you first.
Most of this book has concerned text. The next generation works with more than words: images, video, audio, all in the same system.
For agents specifically, this matters most as an input. An agent that can read a photograph of a damaged product, a scanned invoice, or a screenshot of an error can act on far more of the world than one restricted to typed text. Ask it what is wrong with the thing in the picture, and it can check your order history and open a return in one pass.
Table 17-1: What to watch, and why
| Trend | Why it matters |
|---|---|
| Teams of agents | More reach, with the costs and controls of Part V |
| Shared standards for tools | Integrations stop being built from scratch every time |
| Tasks that run for hours or days | Durable memory and resumable execution stop being optional |
| Agents running on your own device | Lower latency, and data that never leaves |
| Specialists by domain | Law, finance, healthcare, engineering |
| Evaluation and verification | Currently the weakest link in the whole field |
The last row is the one I would watch most closely. We can build agents considerably faster than we can confirm that they work. Whoever makes evaluation straightforward will unlock more than the next capability jump does — which is why Chapter 16 exists.
Want to build something? A route that works.
Use the tools first. Spend real time with Claude, ChatGPT or Gemini. Notice when they search, when they use a tool, when they get it wrong. Intuition comes from use, and it is not replaceable by reading.
Build one agent with one tool. Not a multi-agent system. One agent, one tool, one task you actually have. Listing 1 in Chapter 10 is about the right size to start from.
Start from a problem, not from the technology. Something that annoys you every week is the ideal first project, because you will know immediately whether it worked.
Add complexity only when forced. Add a second tool when the first is not enough. Add memory when starting from nothing every time begins to hurt. Add a second agent when you can say what it is buying you. Every addition should be a response to a problem you actually hit.
Put the approval where the consequence is. Before you build anything that spends money or sends a message, decide who confirms it. That decision is easier now than later.
Measure from the beginning. Twenty test cases and a note of what each run cost. This is the habit that separates a demonstration from a system.