AI Agents

An agent is a tool that takes actions rather than returning text, and that one difference changes everything about how you should evaluate it. A chatbot that gives a wrong answer wastes a minute. An agent that takes a wrong action books the wrong flight, emails the wrong list, or deletes the wrong rows. The tools worth using are the ones that have thought carefully about that asymmetry.

The first question to ask of anything in this subcategory is where it stops. Good agents have explicit checkpoints: they will research, plan and draft freely, then pause before anything irreversible — spending money, sending a message to a real person, writing to a production system. Agents that run the whole chain unattended demo beautifully and are difficult to trust with anything that matters. The launches here that have stayed useful past their launch week are almost uniformly the ones that ask.

The second question is what the agent can actually reach. An agent with no integrations is a chat window with extra steps. An agent wired into your calendar, your inbox, your repository or your browser can do real work, and the same wiring is what makes the permissions question serious. Read what scopes it requests. A tool that asks for full mailbox access to schedule meetings is asking for more than it needs.

Reliability in this space is unusually hard to judge from a landing page, because the failure mode is not a crash — it is a plausible action that was wrong. The most informative thing a maker can publish is a description of what happens when a step fails midway: does it roll back, does it stop and report, or does it barrel on with a broken assumption? Tools that publish this are telling you they have hit the problem and dealt with it.

Cost behaves differently here too. Agents make many model calls per task, so a job that costs a fraction of a cent to answer as a chat message can cost meaningfully more as an agent run. If pricing is per-task or per-run, ask what a run means; if it is token pass-through, expect the bill to scale with how ambitious your instructions are.

Local and self-hosted options exist and are growing, particularly for agents that touch code or private documents. They are slower and require setup, and they never send your repository anywhere. For a broader look at how these launches are evaluated, see [where to launch an AI product](/blog/launch-platforms-for-ai-products).