Near-term agents get more hands.

Parts 1–9 mapped chat → reasoning → tools → loop → risk → fit → policy → eval → ship. This last part names patterns already peeking through. None of them require science-fiction timelines. All of them still need the same recommendation: goals, tools, checkpoints.

Multi-agent systems

Instead of one generalist loop, specialists hand off: researcher → writer → reviewer. The mechanism is still tool calling plus a router. The new failure mode is a bad handoff — the reviewer never sees the tool error the researcher swallowed. Eval the chain, not only each persona’s last paragraph.

Agentic RAG

Retrieval-augmented generation (Lewis et al., 2020) was “fetch documents, then write.” Agentic RAG is retrieve from your approved knowledge, then act inside that fence: open the ticket the doc describes, fill the form the policy allows. To safeguard data, retrieval needs ACLs (Access Control Lists). Retrieval plus tools is an agent with a library card, which is better than an agent with the whole internet — if the corpus is actually yours and actually current.

Computer-use agents

APIs are the sane default. Computer-use is the escape hatch when there is no API: the model looks at screenshots and returns clicks and keystrokes. OpenAI’s computer-use API guide describes that loop: screenshot → action → new screenshot. Anthropic’s computer use tool is the same shape from another lab. Permissions get harder, not easier. A click is a write to whatever is on screen. Human confirmation belongs on anything that looks like purchase, send, or admin.

AI-native applications

Products built around a goal and a toolbelt, not a chat box bolted onto last year’s UI. The user states an outcome. The app runs a constrained agent. Chat is a log, not the product.

Autonomous software development

Longer coding loops with tests as the scoreboard — agents that open branches, run CI, and ask for review. Taste, product risk, and “should we build this?” stay human. Autonomy here means fewer keystrokes, not fewer owners.

Some products worth mentioned

The patterns above already have product instances.

OpenClaw is a self-hosted gateway: one process that holds sessions, tools, memory, and channel plugins — chat apps in, agent out. That is an AI-native app with a toolbelt you run yourself, not a new layer in the stack.

Claude’s computer-use tool is Anthropic’s spelling of screenshot → click → new screenshot. OpenAI’s agents guide is one SDK/runtime spelling of a constrained loop. Neither replaces the gates from Parts 7–9.

What not to wait for

A single agent that “does the company.” A guarantee that computer-use is safe because the demo was smooth. A future where evals become optional.

The spine from Parts 7–9 does not expire when the pattern names change.

Conclusion

We made AI good at answers. Then we let it think longer. Then we gave it tools. The next phase is more of that loop — more agents, more retrieval, more UIs — only if we keep a human hand on the irreversible steps.

If you read one habit into this quarter, make it this: stop asking only “What should I prompt?” Start asking “What may it touch, what must I approve, and how will I see what it did?”

Takeaway: What’s next is still the agent loop — multi-agent, retrieve-then-act, computer-use, named products included — with the same gates you already designed.

Sources

Part 9: Shipping and Deployment
Series start: The Chatbot Era
Next series: How AI Agents Actually Work