Agents and automation — what is actually possible today
The shift from answering to doing
For the first three years of mainstream AI, the model was simple: you ask, it answers. You prompt, it generates. The human takes the output and does something with it.
AI agents change this. An agent is an AI system that doesn’t just produce text — it takes actions. It can search the web, run code, fill out forms, send emails, and chain multiple steps together to complete a task that would otherwise require human intervention at each step.
This is genuinely new. And it’s happening now.
What agents can do today
Multi-step research: An agent given “research the top 5 competitors of Company X and produce a comparison table” can search the web multiple times, read pages, extract information, and compile a structured output — without you doing anything between the first prompt and the final result.
Code execution: Agents can write code and then run it. This means they can process data, generate charts, manipulate files, and do calculations — not just describe how to do these things.
Computer use: Some agents (Claude’s computer use, OpenAI’s Operator) can literally control a computer — clicking buttons, filling forms, navigating websites, taking actions in software. This is early and imperfect but real.
Workflow automation: Agents can be connected to tools like email, calendars, and project management software to take actions on your behalf — scheduling meetings, sending updates, creating tasks.
A standard AI conversation is like asking someone for advice. An agent is like asking someone to actually do something. The difference sounds small but the implications are significant — both for what's possible and for what can go wrong.
What agents cannot do reliably yet
Agents are impressive in demos and genuinely useful for specific tasks. They’re not yet reliable enough to be trusted for consequential tasks without supervision.
Common failure modes: getting stuck in loops, taking wrong actions confidently, misinterpreting ambiguous instructions, failing when unexpected situations arise, and occasionally doing something completely different from what was intended.
The current best practice: use agents for tasks where mistakes are recoverable and you can review outputs. Don’t use them for anything where an error has serious consequences — financial transactions, communications you can’t recall, decisions that can’t be undone.
Practical ways to use agents today
The most reliable uses of agents today are in contained, reversible contexts:
Research and summarisation: Ask an agent to research a topic, read multiple sources, and produce a summary. You review before using.
Data processing: Give an agent a dataset and ask it to clean, transform, or analyse it. Review the output.
Draft generation at scale: An agent that drafts a week of social posts, email responses, or reports — and you review and edit before anything is sent.
Internal tools: Many companies are deploying agents for internal tasks where errors are caught before reaching customers.
The near future
Agent capabilities are improving faster than almost any other area of AI. Tasks that require extensive human supervision today will be more reliably autonomous in 12-24 months. The pattern of AI development suggests: learn to work with agents now in low-stakes contexts, so you’re ready when they become reliable enough for higher-stakes use.
Open Claude or ChatGPT and give it a multi-step task that requires it to reason through several stages: "Research [a company or topic you're interested in], identify the 3 most interesting recent developments, and write a one-paragraph summary of each with your assessment of why it matters." Watch how it works through the task. Notice where it does well and where it's uncertain. This gives you a ground-level feel for current agent capability.